A photoacoustic patch for three-dimensional imaging of hemoglobin and core temperature

Electronic patches, based on various mechanisms, allow continuous and noninvasive monitoring of biomolecules on the skin surface. However, to date, such devices are unable to sense biomolecules in deep tissues, which have a stronger and faster correlation with the human physiological status than those on the skin surface. Here, we demonstrate a photoacoustic patch for three-dimensional (3D) mapping of hemoglobin in deep tissues. This photoacoustic patch integrates an array of ultrasonic transducers and vertical-cavity surface-emitting laser (VCSEL) diodes on a common soft substrate. The high-power VCSEL diodes can generate laser pulses that penetrate >2 cm into biological tissues and activate hemoglobin molecules to generate acoustic waves, which can be collected by the transducers for 3D imaging of the hemoglobin with a high spatial resolution. Additionally, the photoacoustic signal amplitude and temperature have a linear relationship, which allows 3D mapping of core temperatures with high accuracy and fast response. With access to biomolecules in deep tissues, this technology adds unprecedented capabilities to wearable electronics and thus holds significant implications for various applications in both basic research and clinical practice.

Electronic patches, based on various mechanisms, allow continuous and noninvasive monitoring of biomolecules on the skin surface. However, to date, such devices are unable to sense biomolecules in deep tissues, which have a stronger and faster correlation with the human physiological status than those on the skin surface. Here, we demonstrate a photoacoustic patch for threedimensional (3D) mapping of hemoglobin in deep tissues. This photoacoustic patch integrates an array of ultrasonic transducers and vertical-cavity surfaceemitting laser (VCSEL) diodes on a common soft substrate. The high-power VCSEL diodes can generate laser pulses that penetrate >2 cm into biological tissues and activate hemoglobin molecules to generate acoustic waves, which can be collected by the transducers for 3D imaging of the hemoglobin with a high spatial resolution. Additionally, the photoacoustic signal amplitude and temperature have a linear relationship, which allows 3D mapping of core temperatures with high accuracy and fast response. With access to biomolecules in deep tissues, this technology adds unprecedented capabilities to wearable electronics and thus holds significant implications for various applications in both basic research and clinical practice.
Monitoring biomolecules in the human body can help track wellness levels, diagnose diseases, and evaluate therapeutic outcomes. In particular, the amount and location of hemoglobin in the body provide critical information about blood perfusion or accumulation in that area. Low blood perfusion inside the body may result in severe organ dysfunctions. It can happen in many kinds of diseases (such as myocardial infarction 1 , post-cardiac arrest syndrome 2 , and vascular diseases of the extremities 3 ), or after surgery (such as organ transplant 4 ). On the contrary, accumulation of blood is often a sign of inflammation 5 , trauma 6 , or cancer 7 . For example, cysts with many possible types of biofluids inside may be found throughout the human body. Bloody cysts are suspicious and should be further examined and closely monitored for the risk of malignant tumors 8,9 . Continuous monitoring can benefit understanding and diagnosing these pathophysiological conditions, and thus enable timely medical interventions to achieve better outcomes. However, existing methods are not designed for continuous monitoring on individual patients: some necessitate costly equipment, such as magnetic resonance imaging; some rely on radioactive tracers, such as positron emission tomography 10 . Ultrasonography can image internal tissues and blood flow, but requires an operator and a separate lasing system for biomolecule sensing 11 . The recent advances in soft electronics have given rise to soft patches that can adhere to the human skin for continuous health monitoring [12][13][14] . These devices have demonstrated their capability in biomolecule sensing based on electrochemical reactions [15][16][17][18][19][20][21][22] and optics 23,24 . However, existing soft patches can only sense biomolecules close to the skin surface. None of them has access to biomolecules in deep tissues, which have a stronger and faster correlation with the physiological and metabolic processes in the human body than those close to the skin surface 25 (Supplementary Note 1, Supplementary Table 1).
Here we report a photoacoustic patch for continuous sensing of biomolecules in deep tissues. The device integrates an array of high-power VCSEL diodes and piezoelectric transducers, which are interconnected by serpentine metal electrodes and encapsulated in an elastomeric matrix. Pulsed laser emitted from the VCSEL array excites hemoglobin molecules to radiate acoustic waves. Those photoacoustic waves will be received by the transducer array and then processed to reconstruct a 3D map of the hemoglobin with a sub-millimeter resolution. Moreover, the photoacoustic signal amplitude has a linear relationship with the media temperature 26 , which provides a noninvasive way for core temperature measurement with a high spatial resolution and fast response. This work integrates laser sources and piezoelectric transducers into an electronic patch, which is unique in design, fabrication, and working principle among all existing wearable electronic patches (Supplementary Note 2).

Results
Design, fabrication, and working principle of the soft photoacoustic patch Figure 1a schematically illustrates the design and working principle of the soft photoacoustic patch. The patch includes a VCSEL array as the light source and a piezoelectric transducer array for photoacoustic wave detection. The laser beams are diffused in deep tissues. Hemoglobin molecules will undergo thermoelastic expansion after absorbing optical energy and collapse when the energy is absent. Therefore, when illuminated by the pulsed laser from the VCSEL array, hemoglobin will vibrate and emit acoustic waves. The piezoelectric transducers will receive the acoustic waves for generating the spatial distribution of the wave emitters. Therefore, photoacoustic imaging takes advantages of the unique absorption characteristics of biomolecules and highly penetrating acoustic waves to achieve high spatial resolution mapping of biomolecules in deep tissues.
In the patch, 24 VCSELs are evenly distributed in four equally spaced columns (Methods, Figs. S1-S3). The VCSELs in each column are connected in series. The distributed VCSEL layout can help generate uniform illumination in regions below the patch (Fig. S4). 240 piezoelectric transducers are arranged in between the VCSELs, in 15 columns with 16 transducers in each column. To address each transducer independently with a compact device profile, seven layers of serpentine interconnects are designed (Fig. S5). Four adjacent elements in the column are virtually connected in parallel to enhance the signal in the image reconstruction process (Fig. S6), forming 13 linear arrays in the row direction (Fig. S7). The overlap between adjacent arrays can increase the number of imaging planes, further improving the lateral resolution in the overlapping direction (Fig.  S8). The VCSELs, transducers, and interconnects are all encapsulated in an elastomeric polymer, forming an "island-bridge" structure with an overall footprint of 2.0 cm × 1.6 cm and a thickness of 1.2 mm (Methods). In comparison with conventional photoacoustic imaging systems that need complicated components and strict operation environments (Fig. S9  2), the integrated wearable photoacoustic patch greatly reduces the physical constraint on the human body, potentially allowing imaging on moving subjects.
High-power VCSELs are used in this study to achieve a great detection depth and a large signal-to-noise ratio (SNR). A wavelength of 850 nm is used because it has deep-tissue penetration and is in the first optical window for probing human tissues 27,28 (Supplementary Note 4). Hemoglobin also has the dominant optical absorption coefficient compared with other molecules, such as water and lipid, at this wavelength. Furthermore, VCSEL at 850 nm wavelength is most commonly available because, on the one hand, 850 nm is a common optical wavelength whose attenuation in fibers is relatively low 29 ; on the other hand, silicon-based 850 nm photodetectors are low-cost and widely used 30 . The receiving transducer element is composed of a piezoelectric layer and a backing layer (Figs. S10, S11). The piezoelectric layer is made of 2 MHz lead zirconate titanate (PZT) micropillars embedded in epoxy. Compared with bulk PZT, the 1-3 composite suppresses the transverse vibration and enhances the axial vibration of the PZT micropillars, thereby increasing the electromechanical coupling coefficient and improving the energy conversion efficiency. The backing layer, made of cured silver epoxy, has a high electrical conductivity and a strong attenuation effect on acoustic waves to dampen excessive vibrations and thus improve the signal bandwidth and axial resolution of the transducers (Fig. S12).
The as-fabricated soft photoacoustic patch is mechanically and electrically robust. Figure 1b shows the photographs of the patch under different modes of deformation, including bending on a developable surface, wrapping on a nondevelopable surface, and twisting and stretching. The photoacoustic patch is rigid locally at each piezoelectric transducer element and laser diode but soft globally on the system level. No external pressure is required to conformally attach the photoacoustic patch to the skin (Fig. S13). Figure 1c presents the infrared camera images of the patch during operation (850 nm laser wavelength). Mechanical deformations do not affect the performance of the VCSELs (Fig. S14).
Optical, thermal, and acoustic characterizations of the soft photoacoustic patch The optical energy distribution in the tissue should be as uniform as possible to minimize systematic artifacts introduced to the photoacoustic images. Optical attenuation needs to be minimal to ensure the greatest detection depth. Figure 2a shows the simulated optical intensity in a 2 cm × 2 cm × 2.5 cm human muscle tissue (Methods). Each VCSEL emits a laser beam perpendicular to the diode surface into the tissue with a divergence angle of 20°and a peak power of 40 W per VCSEL (Methods). The top surface in Fig. 2a corresponds to the interface between the patch and the tissue. The simulated optical intensity distribution in four planes cross-sectioning the illuminated volume is characterized (the bottom panels of Fig. 2a). The origin of coordinates is set at the central point of the photoacoustic patch. The optical intensities in the plane 1 (XZ plane at y = 0) and plane 2 (YZ plane at x = 0) show uniform distribution in the horizontal direction and small attenuation along the axial direction. In the XY plane, the distributions are highly uniform, with 20% and 2.3% of the incident intensity at the depths of 1 cm (plane 3) and 2 cm (plane 4), respectively. This indicates the ability of the laser beams to penetrate thick tissue layers. The optical intensity distributions of a stretched, bent, and twisted VCSEL array under normal mechanical deformations induced by the skin are tested. The 20% uniaxially stretched array shows an optical distribution very similar to that of an undeformed array (Fig. S15, Methods).
VCSELs in operation will generate a lot of heat. Excessive heat will not only raise safety concerns, but also degrade the VCSEL performance 31 and change the sensitivity of piezoelectric transducers (Figs. S16, S17). Figure 2b shows the thermal images of the patch on a human arm immediately after turning on the lasers (top panel) and after continuous operation for an hour (bottom panel) at a repetition frequency of 3 kHz and a pulse duration of 200 ns (Fig. S18). A relatively long pulse, i.e., 200 ns, is used to enhance the signal-to-noise ratio, which is close to the pulse duration in other studies that utilize LEDs or laser diodes as the light sources [32][33][34] (Fig. S19). After one-hour operation, the maximum temperature measured was~36°C, slightly higher than the skin surface temperature, but still comfortable for the subject.
The photoacoustic patch generates as much heat as ultrasound-phased arrays 35 , both of which are within safety standards. Figure 2c presents the changes in incident intensity from the VCSELs during continuous operation. At 3 kHz pulse repetition frequency and 200 ns pulse duration, the intensity decreases by only <4% after one hour, showing the high stability of the VCSELs.
For quantitative photoacoustic studies, it is critical for the transducer array to have a uniform distribution of detection sensitivity Fig. 1 | Design, fabrication, and working principle of the soft photoacoustic patch. a Schematics of the device structure and the working principle. The patch comprises an array of VCSELs and an array of piezoelectric transducers, interconnected by serpentine copper electrodes. All components are encapsulated in Ecoflex. Upon absorption of the optical energy, the hemoglobin molecules in red blood cells undergo thermoelastic expansion and radiate acoustic waves into the surrounding media. The photoacoustic waves will be collected by the transducer array and then relayed to a backend system for data processing. AlN: aluminum nitride. VIA: vertical interconnect access. VCSEL: vertical-cavity surface-emitting laser. b Optical photographs of the soft photoacoustic patch under different modes of deformation, including bending on a developable surface, wrapping on a nondevelopable surface, and stretching and twisting. Insets in the middle and right panels are optical micrographs of a single transducer element and a VCSEL diode, respectively. c Infrared camera images of the soft photoacoustic patch when the VCSELs (850 nm laser wavelength) are in operation under different modes of deformation, including stretching, bending, and twisting. to photoacoustic signals in the target region. Figure 2d shows the simulated photoacoustic sensitivity distribution of the patch in the 2 cm × 2 cm × 2.5 cm human breast tissue in consideration of nonuniform light distribution in Fig. 2a (Methods, Fig. S20). Due to the remarkable penetration of acoustic waves in human tissues, the detection sensitivity loss caused by the wave-sensing ability of piezoelectric transducers is < −10 dB. The high detection sensitivity ensures the high imaging depth of the photoacoustic patch. The tested transmitting and receiving properties of piezoelectric transducers demonstrated their high penetration depth and uniform sensitivity (Figs. S21-S23, Methods).
The impulse response is a critical characteristic of a sensing system (Methods, Figs. S24, S25), which is characterized by the time domain photoacoustic signal of a linear source excited by the VCSELs in this study. We measured the photoacoustic signals of a human hair (with a diameter of~80 µm). The working frequency of the system is then characterized by applying Fourier Transform to the temporal photoacoustic signal (black curve) received by one transducer element (Fig. 2e). As the optical intensity of VCSELs is much lower than a conventional bulky laser, photoacoustic signals in the time domain are averaged to increase the SNR. On the other hand, the times of averaging will reduce the frame rate of imaging. To balance the SNR and  frame rate in this study, the times of averaging are 3000 (Fig. S26), yielding a frame rate of 1 Hz and a SNR 36-38 of 26.8 dB (signal of a hair at a depth of 2 cm in a gelatin phantom) at a lasing pulse repetition frequency of 3 kHz. The red curve shows the impulse response in the frequency domain, with a center frequency of 2.40 MHz and a bandwidth of 1.47 MHz (Figure 2e). Imaging resolutions are characterized based on a linear source. Photoacoustic images are reconstructed based on signals generated by hairs embedded in gelatin phantoms at different depths [39][40][41][42] . Figure 2f displays a 2D photoacoustic image of a hair at a depth of 2 cm (Methods, Fig. S27). The amplitude profiles (black dots) of the photoacoustic image in the lateral and axial directions are fitted by the Gaussian function, illustrated by the blue and red curves, respectively (the bottom panel). The image resolution is determined by the full width at half maximum (FWHM) of the Gaussian curve fit to the profile [40][41][42][43] . The axial resolution is mainly determined by the signal frequency and remains almost constant at~0.8 mm for different imaging depths (Fig. 2g). The lateral resolution will be reduced from 0.4 mm to~0.7 mm as the imaging depth increases because of the degraded focusing (Fig. 2g).

Ex-vivo 3D hemoglobin mapping and core temperature measurement
The wavelength of 850 nm is critical for a high penetration depth in human tissues 27,28 . Additionally, for photoacoustic mapping of hemoglobin amongst other biomolecules in the tissue, a laser wavelength where hemoglobin absorption is dominant needs to be selected. To characterize the sensing selectivity at this wavelength, we tested cyst phantoms with five different biofluid inclusions, including water, plasma, milk, fat, and bovine whole blood, in transparent colorless silicone tubes embedded underneath a 2 cm thick porcine tissue (Fig.  S28). Figure 3a shows the measured optical absorption spectra of all types of biofluids (Methods), which shows that bovine whole blood has the dominant absorption coefficient at 850 nm. To further verify the selectivity, both ultrasound and photoacoustic images of the cyst phantoms are collected (Fig. 3b, Methods). Ultrasound-based B-mode images can only detect the acoustic impedance mismatch between different tissues, which is why the boundaries between the inclusions and matrixes are clear, while the types of biofluids are indistinguishable. The photoacoustic images are based on the contrast of optical absorptions, which differentiates the blood from other biofluids (Fig. S29).
16 rows of transducers form 13 linear arrays, each of which can produce a 2D photoacoustic image. Combining the 13 images, the patch can generate a 3D map of hemoglobin. The 3D mapping performance is tested on two crossed silicone tubes filled with bovine blood embedded underneath a 2 cm thick porcine tissue. Figure 3c shows a slice of the 3D map where the two silicone tubes overlap (top panel) and the corresponding photoacoustic signal amplitude profile along the green dashed line in the top panel (bottom panel). All 13 slices of the photoacoustic images are displayed in Fig. 3d, where the slice with overlapped tubes in Fig. 3c is highlighted. Figure 3e gives the integrated 3D mapping of hemoglobin at a depth of 2 cm (Figs. S30, S31). The patch can also distinguish two overlapping vessels (Fig. S32). When the stretching strain is within 15%, the patch performance is minimally affected (Fig. S33).
Core temperature is critical for governing the essential functions of the body and should be maintained near 37°C 44 . It typically fluctuates within 1°C according to circadian rhythm 45 , but can reach 40°C amid strenuous workload or~35.6°C in cold environments. A significant deviation of the core temperature indicates failing thermoregulation 44 with dire consequences [46][47][48][49] , sometimes lifethreatening. Most soft patches can only measure the temperature on the skin surface, which can be easily affected by the external environment and thus has a weak correlation to the core temperature.
Photoacoustic signals are generated when the biomolecules convert the pulsed optical energy to mechanical energy in the form of photoacoustic waves 54 . In the range of 10-55°C, there is a linear relationship between the amplitude of photoacoustic waves and the temperature 26 (Supplementary Note 6), allowing the measurement of temperature by the photoacoustic approach. As an initial test, we used the soft photoacoustic patch to measure the temperature in a phantom and checked its performance with thermocouples (Fig. S34). The phantom is composed of warm bovine blood injected in transparent silicone tubes underneath 2 cm thick room-temperature porcine tissues. Thermocouples were placed in the tubes, where the photoacoustic measurements were also taken. We demonstrate the high accuracy (Fig. S35, Supplementary Note 7), spatial temperature mapping, and fast response (Fig. S36) of the photoacoustic patch in core temperature measurement by detecting static blood, as validated by the thermocouple.
To monitor flowing blood, we tested the core temperature of an ex-vivo porcine tissue (Fig. S37). A pump drove the blood to flow in a transparent silicone tube with an inner diameter of 3 mm. The flow rate was set to be~9 mL s −1 , resulting in a blood flowing speed of 127 cm s −1 , faster than the blood flow velocity of most blood vessels in the human body 55 . The two ends of the tube were immersed in a beaker containing bovine blood, which was placed on a hot plate to heat the blood to different temperatures during flowing. The blood could also naturally cool down while the hot plate was turned off. A portion of the tube was embedded underneath a porcine tissue at a depth of~2 cm, which was measured by the photoacoustic patch. A thermocouple was inserted into the tube to measure the blood temperature simultaneously and record the data continuously. Figure 3f shows the beamformed photoacoustic amplitude (black dots) of flowing blood measured by the photoacoustic patch and the temperature (red line) measured by the thermocouple, which agree well with each other during the entire dynamic process. Two cycles of heating and cooling were tested, lasting~82 min. Figure 3g presents the scatter plot of the photoacoustic amplitude as a function of the thermometer temperature. The fitting of the measured data (R 2 = 0.987) demonstrates the linear relationship between the photoacoustic amplitude and the temperature in flowing blood.

In-vivo 3D imaging of blood vessels and venous occlusion test
To test the feasibility of in-vivo monitoring, we used the photoacoustic patch to image veins in the hand, foot, thigh, and forearm, monitor the venous response to the occlusion test, and image the internal jugular vein (IJV) (Fig. S38, Methods). Figure 4a-h presents the photos of a volunteer's hand, foot, thigh, and forearm, where the target veins are labeled. The photoacoustic patch acquired 13 cross sections of the veins, which were then converted to 3D images, respectively, as shown at the sides of corresponding photos. The 3D images clearly display the vein structures of different body locations. In comparison to Doppler ultrasound imaging 35 , photoacoustic imaging has high sensitivity and contrast in detecting blood vessels, especially for small blood vessels with slow blood flow 56 (Fig. S39). Note that the flexible photoacoustic patch may also be affected by motion artifacts (Fig. S40), a common problem in existing wearable electronic devices [57][58][59] . The volunteer kept the arm static during the measurements to minimize any motion artifacts.
Venous occlusion plethysmography is a noninvasive tool to assess the blood flow and vascular resistance of limbs 60,61 . In the measurements, venous return from the forearm was briefly interrupted by inflating a cuff, wrapped around the upper arm, to above venous pressure but lower than the diastolic pressure (Methods).
As a result, the venous dimension will increase as the arterial blood inflow. We attached the photoacoustic patch on the forearm, above the veins, and continuously monitored the dynamic vascular response to a venous occlusion (Supplementary Movie 1). Figure 4i shows the change of the vein size during a 3 min continuous recording. Insets show the photoacoustic images of the vein at three different moments. The image pixels with a normalized value >0.5 were counted into the vein area 62 , as labeled by the white boundaries. No pressure was applied in the first 1 min, thus no significant area change was observed. At 60 seconds, the cuff was quickly inflated to 70 mmHg, resulting in an increasing vein area with the time. Inset image 2 presents an obvious expansion of the vein compared to inset 1. The cuff was rapidly released after 1-min inflation, accompanied by a dramatic drop in the vein area. Those results are similar to others acquired by bulky photoacoustic systems 63 . The venous occlusion test demonstrated the fast response of the photoacoustic patch for in-vivo imaging.
We used the photoacoustic patch to 3D image the IJV (>1.1 cm in depth) in the neck ( Fig. 4j and Fig. S41). Figure 4k shows 13 slices of photoacoustic images of the IJV superimposed on the corresponding b Ultrasound B-mode and photoacoustic images of different cysts embedded in a porcine tissue at a depth of 2 cm. The ultrasound B-mode images, acquired by a commercial ultrasound probe, show no differences because of the low acoustic impedance contrast between various fluids. The soft photoacoustic patch differentiates the blood cyst based on the high optical absorption contrast. c A high-resolution photoacoustic image of a blood vessel phantom, which contains two silicone tubes filled with blood embedded underneath a 2 cm thick porcine tissue. The red curve at the bottom panel shows the profile along the green dashed line. d 13 slices of photoacoustic images of the blood vessel phantom with display transparency of 80%. e 3D imaging of hemoglobin at a depth of 2 cm in porcine tissue. f Comparison between the beamformed photoacoustic signal amplitude and temperature of flowing blood during two cycles of heating and cooling. g Linear fitting of the photoacoustic amplitude and temperature. ultrasound B-mode images (Fig. S42), which are all acquired by the photoacoustic patch. The central frequency of the ultrasound transducers is close to 2 MHz, which results in the low contrast of the IJV in the ultrasound B-mode image. On the contrary, the IJV shows high contrast to other surrounding tissues in the photoacoustic image (Fig.  S43), benefiting from the strong optical contrast between the hemoglobin and other molecules. The irregular skin curvature has a minimal influence on the imaging resolutions of the photoacoustic patch, due to the relatively low acoustic working frequency (Figs. S44, S45 and Supplementary Note 8). The carotid artery is invisible in the photoacoustic images because its strong pulsation will induce unstable phases to the photoacoustic signals and therefore damage their coherent averaging 64,65 (Fig. S46).

Discussion
The soft photoacoustic patch demonstrated in this study allows for continuous, noninvasive mapping of hemoglobin and core temperature with high spatial resolution in real time. This work reports using soft electronic devices for 3D imaging of biomolecules in deep tissues (>2 cm in ex-vivo tests and >1.1 cm in in-vivo tests). The high-resolution imaging of hemoglobin will enable the monitoring of hemodynamics and vascular proliferation in tissues to manage a variety of conditions and diseases. Monitoring the dimension of blood vessels can be valuable for evaluating vessel functions and diagnosing vascular diseases. For instance, measuring the dynamic change of the vein diameter during an occlusion can help examine venous compliance, which is a strong indicator of cardiac function 60 . The photoacoustic effect-based are counted to calculate the vein size. j A photograph of the patch attached to the human neck at a location above the internal jugular vein. k 13 slices of dual-mode images acquired by the photoacoustic patch, i.e., photoacoustic images of the internal jugular vein superimposed on ultrasound B-mode images. The pixel values of normalized photoacoustic images < 0.5 are not shown. temperature measurements, with the advantages of deep penetration, high accuracy, and fast response, introduce a strategy for monitoring the core temperature, e.g., during exercise, anesthesia, and surgical hypothermia, in fundamental biomedical research and clinical practice 66,67 .
Although the photoacoustic patch discussed here only detects hemoglobin, this platform technology can potentially be extended to monitor many other endogenous biomolecules, such as melanin 68,69 , glucose [70][71][72] , lipid 73,74 , cytochrome 75 , nucleic acid 76 , and proteins 77,78 (Supplementary Note 9). Furthermore, exogenous contrast agents, like single-walled carbon nanotubes 79 , gold nanoparticles 80 , and methylene blue 81 , can further enhance the signal intensity, increase the detection depth, and improve the detection specificity 82 . The laser wavelength is the key to selectively monitoring various biomolecules. Integrating multiple laser diodes with different wavelengths on the photoacoustic patch can expand the portfolio of detectable biomolecules, with more accurate targeting of biomolecules by detecting a set of absorption characteristics at different wavelengths. It is possible to integrate two or more wavelengths VCSELs in the patch as VCSELs with a wide wavelength range have been developed, from blue light 83 (447 nm) to infrared light 84 (1550 nm), which overlaps with the typical wavelength range in photoacoustic imaging applications.
The current detection depth is still limited by the optical intensity of the VCSELs. It is challenging for the photoacoustic patch to image the cardiac region because the optical power of laser diodes used in this work is not as high as expensive bulky high-power lasers. The depth of the cardiac region ranges from several centimeters to >17 centimeters 85 . The average shortest distance between the heart and the skin is about 3.1 cm when detecting from the apical view 86 . This large imaging depth is currently not achievable by the wearable photoacoustic patch. In-vivo photoacoustic imaging of the human heart represents a grand challenge in the field, even using conventional bulky photoacoustic systems with expensive high-power lasers. In this work, we achieved an ex-vivo imaging depth of~2 cm in porcine tissues and an in-vivo imaging depth of~1 cm. Higher power VCSELs will be needed to further increase the detection depth to the regions of visceral organs. Additionally, developing higher power VCSELs, by either fabricating larger VCSELs with more light emitting elements (Figs. S3, S47) or constructing driving circuits with higher output current, will be essential for increasing the SNR of photoacoustic signals and thus reduce the times of averaging for imaging dynamic arteries.
Photoacoustic imaging without calibration can only monitor relative temperature changes. A calibration process that establishes the relationship between photoacoustic amplitude and absolute temperature can enable photoacoustic imaging to monitor absolute temperature. In this work, we demonstrated monitoring absolute core temperature in ex-vivo porcine tissues. While monitoring the photoacoustic amplitudes of blood, we recorded the absolute temperature simultaneously by a thermocouple. Therefore, the photoacoustic amplitude could be calibrated and transferred to absolute temperature. For in-vivo applications, photoacoustic imaging can monitor relative temperature changing without calibration. Invasive temperature catheters can be used for calibration. It is worth mentioning that for long-term monitoring, the calibration is only required once.
Before applying the photoacoustic patch to monitor blood temperature in the human body, additional challenges need to be solved. First, a gold standard technique for core temperature measurement is required to calibrate the photoacoustic amplitude. Some noninvasive core temperature sensors based on thermal flux models suffer from slow response and lack of spatial resolution, which are not suitable for calibration in this case. Invasive catheters can be directly inserted into the blood vessel to monitor temperature, but it is too invasive. Second, some other factors may also affect the photoacoustic signals, such as the amount of blood perfusion. More advanced methods should be developed to eliminate the influence of these factors, such as the thermal memory based photoacoustic technique 87 .
We adopted data averaging and bandpass filter to improve the signal-to-noise ratio. A high pulse repetition rate, i.e., 3 kHz, is used to ensure a high imaging frame rate (Supplementary Note 10). Although data averaging is a common and convenient way, thousands of times of averaging makes it time-consuming, causing more laser exposure. Besides, there are other methods for enhancing the SNR while requiring less time consumption (Supplementary Note 11), such as coded excitation [88][89][90][91] , empirical mode decomposition 92,93 , wavelet thresholding 94,95 , Wiener deconvolution 96 , and adaptive filtering 97 . These methods and the data averaging can be adopted together to achieve better SNR 92 . In the current design, bulky ultrasound probes and sophisticated laser machines are eliminated, which have significantly improved the device portability and ease of use, but the photoacoustic patch is still wired to a backend system for signal acquisition and processing. Future efforts can focus on miniaturizing the control electronics to realize a fully integrated wearable system and, therefore, enable measurements on-the-go (Fig. S48).

Human experiment protocols
All human tests were performed under University of California San Diego Institutional Review Board (IRB) approval (number 800975). One 25-year-old male volunteered to be tested with informed consent obtained without compensation. For the venous occlusion experiment, the volunteer sat on a chair with a pressure cuff worn on the upper arm. The vertical distance between the neck and the forearm was about 30 cm. Then we attached the photoacoustic patch on the forearm above the veins using a medical tape. After that, the venous occlusion was performed: (1) No pressure was applied to the cuff in the first 1 minute; (2) inflate the cuff to 70 mmHg immediately and maintain for 60 s; (3) deflate the cuff to zero to let the veins recover to the normal status. In the detection of the internal jugular vein, the volunteer sat on a chair with the photoacoustic patch attached to the neck with a medical tape. For the imaging of veins in the hand, foot, thigh, forearm and venous occlusion test, a 1 cm-thick gelatin phantom was placed between the patch and skin to compensate non-uniform light distribution (Fig. S38).

Fabrication of laser diode chips
The fabrication process of the VCSEL diode chip is schematically illustrated in Fig. S2. The anode and cathode of the VCSEL die (850 nm, Ace Photonics) are on the top and bottom surfaces (Fig. S3), respectively. To facilitate the fabrication of the photoacoustic patch, the anode and cathode are routed to the same surface by creating vertical interconnect accesses (VIAs) and wire bonding. Two vertical openings were created by laser ablation in a 1.7 mm × 2.4 mm × 0.25 mm aluminum nitride (AlN) substrate and filled with silver epoxy (E-Solder 3022). The silver epoxy VIAs were cured in an oven at 80°C for 2 h. The AlN substrate was cleaned with acetone and isopropyl alcohol to remove organic contaminants, followed by rinsing with DI water and drying with nitrogen gas. Moisture induced in the cleaning process was removed by baking the samples in a vacuum oven at 100°C for 10 min. A lift-off process allowed patterning metal electrodes on AIN. The process involved photolithography (photoresist AZ 1529: spin-casting at 4000 r.p.m. for 60 s, baking on a hotplate at 95°C for 120 s, UV irradiance at 350 mJ cm −2 , and developing for~40 s with developer AZ 300 MIF) and then sputtering ( Moisture induced in the lift-off process was removed by baking the samples in a vacuum oven at 100°C for 10 min. The VCSEL die was then pasted on the ground electrode pad on AlN with silver epoxy, which was cured in an oven at 80°C for 2 h. The anodes of the VCSEL die and AlN substrate were connected with wire bonding.

Fabrication of the photoacoustic patch
The fabrication process can be generalized into three steps: (1) patterning of the stretchable multilayered electrodes, (2) preparation of the VCSEL diode chips and ultrasonic transducer array, and (3) soft packaging. Cu foils with 20 μm thickness were used as the multilayered conductive interconnects. To adhere the interconnects on the soft elastomeric substrate tightly, a PI thin film [poly(pyromellitic dianhydride-co-4,40-oxydianiline) amic acid solution, PI2545 precursor, HD MicroSystems] was spin-coated on the Cu, at the speed of 4000 r.p.m, with an acceleration of 5000 r.p.m per second, for 60 s. The PI was cured by soft baking at 100°C for 3 min and hard baking at 300°C for 1 h under a nitrogen atmosphere. The PI-based Cu foil was activated by ultraviolet light (PSD series Digital UV Ozone System, Novascan) for 2 min and then laminated on a temporary PDMS substrate (base to hardener ratio is 20:1, Sylgard 184 silicone elastomer). The ultraviolet light activation strengthens the bonding between the PI and the PDMS substrate. A nanosecond laser (Laser Mark's, central wavelength, 1059-1065 nm; power, 0.228 mJ; frequency, 35 kHz; speed, 300 mm s −1 ; and pulse width, 500 ns) was used to ablate the Cu/PI into the "islandbridge" serpentine layout. The electrode patterns were designed by AutoCAD (Autodesk, USA). The patterned Cu/PI thin film was transferprinted to an Ecoflex substrate (15 μm thick; Ecoflex-0030, Smooth-On) on a glass slide using a water-soluble tape (3 M) after activation by ultraviolet light for 3 min. To tightly stack the second layer of the electrode on top of the first layer, a dielectric layer (15 μm) of Ecoflex was spin-coated on the first layer. Using the same method, six layers of top stimulation electrodes were built up and aligned under the microscope. The VIAs were developed by laser ablation to route all electrodes in multiple layers to the same plane. The VCSEL array was bonded with the six-layer electrode using silver epoxy (Esolder 3022, EIS, USA). Anisotropic conductive films (Elform) were hot pressed to the front pads of the electrodes to connect the patch to the external power supply and the data acquisition system. The bottom common ground electrode was fabricated in a similar way to the top electrodes.
The structure of the ultrasonic transducer consists of a piezoelectric material and a backing layer. 1-3 PZT-5A composites (Del Piezo, USA) were selected due to their high electromechanical coupling coefficients. The condensed backing layer was made of silver epoxy (Esolder 3022, EIS, USA) for absorbing the extra ultrasonic waves. The silver epoxy composite was mixed with the hardener in a 12.5:1 ratio over 10 min and mounted on a 0.3 mm thick mold, which was then cured at 80°C for 2 h. The same silver epoxy was used to integrate the backing layer with the 1-3 composite material, and the entire piece was diced into multiple small elements (0.8 mm length × 0.6 mm width × 1 mm thickness).
A scaffold with 240 openings was customized to fix the piezoelectric transducers. Connections to the top and bottom electrodes were achieved with the conductive adhesive at 80°C for 2 h. The device was encapsulated by filling the device with the uncured Ecoflex precursor, followed by curing at 80°C for 20 min. After that, the glass substrates carrying the top and bottom electrodes were peeled off. VCSEL chips and piezoelectric transducers are connected to external driving and signal acquisition systems with wires. The connection of VCSEL chips can be integrated with that of the piezoelectric transducers, which does not increase the complexity of the overall wearable patch compared to ultrasound sensors 98 .

Simulation of optical distribution
The simulation of the optical intensity distribution in a 3D space was performed by the Monte Carlo method using an open-source MATLAB toolbox-MCmatlab 99 . A 4 cm × 4 cm × 4 cm homogeneous region was set as the human breast tissue, with the absorption coefficient μ a , scattering coefficient μ s , Henyey -Greenstein scattering anisotropy factor g, and refractive index n set as 0.1 cm −1 , 85 cm −1 ref. 100, 0.9 ref. 99, and 1.3, respectively. The region above the top surface was considered as air, with μ a , μ s , g, and n set as 1 × 10 −8 cm −1 , 1 × 10 −8 cm −1 , 1, and 1, respectively. The laser diode array was placed at the center of the top surface. The width of each laser source was 1.5 mm. Each laser diode emitted a laser beam into the tissue perpendicular to the surface with a divergence angle of 20°. All the boundaries were set to be cuboid. The wavelength was 850 nm.

Simulation of photoacoustic detection sensitivity
The simulation of photoacoustic detection sensitivity was performed in a 4 cm × 4 cm × 4 cm homogeneous region using an open-source MATLAB toolboxk-Wave 101 . The transducer array was placed at the center of the top surface. Assuming the background tissue as the human breast, the sound speed and tissue density were set as 1510 m s −1 and 1020 kg m −3 , respectively. The frequency dependent acoustic absorption coefficient was considered as 0.75 dB (MHz y cm) −1 , where y equals to 1.5 ref. 102. The simulation region was divided into voxel elements with a pitch of 0.05 mm in each direction. In each voxel, one point source emitted a pulsed photoacoustic signal with the amplitude decided by the light distribution. All transducers received the pulse signal, followed by Delay-And-Sum beamforming. The amplitude of the beamformed signal was the detection sensitivity of this voxel.

Characterization of VCSEL array
The laser power of a single VCSEL chip is about 40 W measured by a power meter (Newport Corporation, 835 Optical Power Meter, 818-SL detector, 883-SL attenuator), which has a sensing aperture of 11.3 mm to cover the entire light beam of one VCSEL. Considering the entire patch with a footprint of 2 cm × 1.6 cm, the average power is about 1.8 × 10 3 W m −2 , which is lower than the safety limit 103 of 3.99 × 10 3 W m −2 . Smaller pulse repetition frequency can be selected to further reduce the power if needed by specific use cases. To detect the light fields of optical beams in different cases, including a single VCSEL, and an undeformed, stretched, bent, or twisted VCSEL array, we scanned a photodetector point by point in free space to measure the optical intensity in a 2D plane. An optical attenuator (Thorlabs, NE60A-B) was fixed on the photodetector (Thorlabs, PDA10A2) to make sure the optical intensity does not exceed the measurement range of the photodetector. The scanning plane was 3 cm away from the VCSEL and VCSEL array. We measured five optical fields with a size of 2 cm × 2 cm and a step size of 1 mm (Fig. S15).

Characterization of piezoelectric transducers
The transmitting sound field of a transducer element was measured using a hydrophone (ONDA, Model no. HNP-0400) in a water tank (Fig. S21). The piezoelectric transducer was excited by a pulse voltage of 100 V. A hydrophone scanning system (ONDA, AIMS III) moved the hydrophone in the 3D space. The peak-to-peak value of the signal measured by the hydrophone was extracted. To test the receiving sensitivity of the piezoelectric transducers, the performances of the photoacoustic patch and the commercial probe P4-2v from Verasonics were compared. P4-2v was chosen because it has a central frequency of about 2.7 MHz, close to the transducers (~2.4 MHz) in the photoacoustic patch. A 100 V pulse was applied to a customized single transducer (0.5 mm × 0.5 mm) to emit ultrasound waves, which was measured by both the commercial probe and the photoacoustic patch (Fig. S22). Signals measured by four elements of the photoacoustic patch were summed, which was the same as the case in the practical applications. To map the receiving sensitivity in 3D space, the single transducer ultrasound source was moved by the scanning system (ONDA, AIMS III) to emit ultrasound waves, which was measured by the photoacoustic patch and beamformed in the Verasonics system. The peak-to-peak value of the beamformed signal was extracted (Fig. S23).

System setup and data collection
Verasonics Vantage 256 worked as the host to control the timing sequence of the whole system and signal acquisition. It has 256 individual signal acquisition channels with built-in low-noise amplifier, programmable gain amplifier and filters. That means each element receives the photoacoustic signal independently. All of the elements can receive the data simultaneously. Signals of four elements will be summed digitally in the MATLAB program to form one element in each virtual linear array. A program was written by MATLAB and run on the Verasonics system, controlling the laser radiation and photoacoustic signal acquisition.
To synchronize the laser emission and signal acquisition, Verasonics exported a 3.3 V LVTTL-compatible trigger signal to the signal generator (Rigol, DG822), which was a 1 μs active low output. The signal generator would be triggered to output a 5 V pulse signal with a duration of 200 ns. The laser driver (PicoLAS, LDP-V 240-100 V3.3) received the output from the signal generator, and immediately provided a 50 A current to drive the laser diodes with a pulse duration of 200 ns. The peak power of each VCSEL was 40 W driven by a 50 A pulse current. After laser illumination, the Verasonics system started the signal acquisition process. The recorded photoacoustic signal was digitized at a sampling frequency of 62.5 MHz and filtered by a bandpass filter with a center frequency of 2.2 MHz and −6 dB bandwidth of 1.2 MHz. To enhance the SNR, photoacoustic signals were averaged 3000 times to reduce the incoherent noise. Verasonics controlled the VCSELs to emit laser beams and transducers to receive signals at a pulse repetition frequency of 3 kHz, resulting in a detection frame rate of 1 Hz. A C-language program was written and called in MATLAB by the host program to reconstruct the 2D images. Reconstructing one 2D image takes about 50 ms, which means 0.65 s is required to reconstruct all of the 13 slices of 2D images. These slices of 2D images can be shown during the measurement in real-time, which reveal information in 3D space. Converting the 2D images into 3D image was manually processed offline in a software (Amira) after saving all of the 2D images. The processing time took <20 s. The conversion from 2D images to 3D image may be processed automatically in MATLAB in the future to save time. The time-domain signals were also saved for offline processing to reconstruct the 3D images.
The human skin and driving electrodes that connect VCSEL diodes are isolated by a 1 mm-thick Eco-flex 00-30 layer. As reported 104 , the leakage current for such a silicon polymer layer with the same thickness is as low as 10 −11 A at an applied electric field of 5 V μm −1 . Since the applied electric field in this study is less than 1 V μm −1 , the leakage current should be smaller than 10 −11 A for the photoacoustic patch, which is very safe. In the ex-vivo temperature measurements (Fig. 3f,  g), to avoid direct illumination on the thermocouples (Omega Engineering Inc., Model no. SC-TT-K-30-36) by the laser beam and getting photoacoustic signals from them, the thermocouples were placed at the peripheral region of the photoacoustic patch. Supplementary Tables 4 and 5 list the detailed information of experimental equipment and material used in this study, respectively.

Image reconstruction algorithms
The Coherence-Factor-weighted-Delay-And-Sum (CFDAS) algorithm was applied to reconstruct photoacoustic images. For the unmodified DAS beamforming algorithm, assuming the photoacoustic signals are measured by a transducer array with M elements, the received signal of each channel is expressed as p m (t). To reconstruct the image I(x, z) at pixel (x, z), the wave propagation time from the pixel to the m-th element is calculated as Δt m . Therefore, the image I(x, z) could be computed through the summation of P M m = 1 p m ð4t m Þ. The CFDAS introduces an adaptive coherence factor as an additional weight to 105 . CFDAS has been demonstrated to improve the image quality 106 (Fig. S27). The reconstructed 2D images were combined in Amira to form 3D images. The gaps between the 2D image slices were smoothed by Amira automatically.

Test of optical absorption spectra
The NIR-UV-Vis measurements were carried out through a PerkinElmer lambda 1050 UV/Vis/NIR Spectrometer. Water absorbance spectrum was measured under 150 mm InGaAs Int. Sphere Absorbance module and the rest were carried out through 3D WB Det. Absorbance Module. Before each measurement, a 100% transmittance (0 absorbance) baseline was auto-zeroed. The water spectrum was denoised through white certified reflectance standard from Labsphere Company while the rest background was calibrated with pure water. The detection cuvette had a transmittance length of 5 mm. The injected beam (Slit width of 2.00 nm) was sourced from the combination of D2 Lamp and Tungsten Lamp with a lamp change at 860.8 nm. The spectra were collected in the wavelength range from 1000 nm to 700 nm with a data interval of 1 nm.

Acquisition of ultrasound B-mode images
The ultrasound B-mode images of cyst phantoms were acquired by the Verasonics Vantage 256 with a L11-5v linear array. The center frequency of the probe was 7.8 MHz. The compounding imaging strategy was applied to reconstruct the images, which transmitted plane waves in 21 directions, received the echoes, and combined them all to form a single image.

Reporting summary
Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article.

Data availability
All data supporting the findings of this study are available within the paper and its Supplementary Information. The data generated in this study are available from a public data repository at https://figshare. com/articles/dataset/PA_patch/21440925.

Competing interests
The authors declare no competing interests.
Correspondence and requests for materials should be addressed to Sheng Xu.
Peer review information Nature Communications thanks Fei Gao, Benjamin Tee and the other, anonymous, reviewer(s) for their contribution to the peer review of this work.
Reprints and permissions information is available at http://www.nature.com/reprints Publisher's note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons license, and indicate if changes were made. The images or other third party material in this article are included in the article's Creative Commons license, unless indicated otherwise in a credit line to the material. If material is not included in the article's Creative Commons license and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this license, visit http://creativecommons.org/ licenses/by/4.0/.

Supplementary Note 1 | Comparison between different biomolecular imaging methods
Conventionally, there are several non-invasive methods for detecting biomolecules with high penetration depth (>10 cm) such as magnetic resonance imaging (MRI) 1,2 and positron emission tomography (PET) 3,4 . MRI leverages the different magnetic properties of various biomolecules to generate images. More specifically, those biomolecules possess multiple relaxation times after receiving radiofrequency pulses. PET is based on the detection of two annihilation photons, which are generated by collisions of the injected radioactive tracers and electrons within the tissue. The tracers will only couple with the target molecules so that high contrast will be achieved. However, the existence of radioactive tracers prevents this technique from long-term use. For these two methods, the associated equipment is too cumbersome and expensive, thus impossible for wearable long-term health monitoring.
Several optical methods are used for biomolecular imaging. The principle of fluorescence imaging is that materials will emit fluorescent light at a specific wavelength after absorbing high energy photons. Different molecules possessing various molecular energy structures result in various fluorescence. Although the spatial resolution is high (~4 μm), this technique is highly limited by its shallow penetration depth (~3 mm) 5 . Optical coherence tomography (OCT) typically uses nearinfrared light for imaging. The backscattered light is measured with an interferometric setup to reconstruct the depth profile of the tissue 6 . But the penetration depth is still limited (~2 mm) 7 .
Photoacoustic imaging involves shining a laser beam onto tissues. After that, the light energy is absorbed by the biomolecules and converted to mechanical vibration energy, i.e., photoacoustic waves. Photoacoustic imaging entails several advantages 8 compared to the aforementioned optical imaging: (1) by illuminating various molecules at different wavelengths, photoacoustic tomography exhibits high contrasts regarding chemical compositions; (2) the spatial information of biomolecules is encoded in the ultrasound waves, which has relatively weak attenuation in biological tissues. Therefore, photoacoustic imaging can achieve high spatial resolution (tens of micrometers) mapping of biomolecules in deep tissues (several centimeters in depth) 9,10 . As of now, existing photoacoustic imaging devices are bulky and cumbersome, not suitable for wearable long-term continuous use.
The emergence of wearable devices such as soft electrochemical electronics 11 and soft optoelectronics 12 makes continuous monitoring possible. Nevertheless, these devices are still limited by the shallow detection depth beneath the skin. Additionally, these devices can only provide measurements at specific locations, lacking spatial resolutions.
The soft photoacoustic patch in this work inherits the merits of photoacoustic imaging, which is biomolecular selectivity and high imaging resolution in deep tissues. Also, the flexible and stretchable configuration allows the device to be conformally attached to the skin, which can potentially enable convenient and continuous measurements on the go. By innovative designs in device layout and advanced data processing methods of the soft photoacoustic patch, we achieved a detection depth >2 cm with high spatial resolution in biological tissues.

Supplementary Note 2 | Innovation of the photoacoustic patch in the fields of soft electronics and photoacoustic imaging
The innovation of this work can be summarized as the following points. In soft electronics, we developed the flexible and stretchable electronic patch that can noninvasively detect and image molecules in deep tissues. Skin-like wearable patches that integrate various kinds of sensors can monitor the health and wellness of the human body. Existing skin-like wearable patches can sense biomolecules in sweat [13][14][15] , saliva 16,17 , and tears 18-20 on the skin surface, or interstitial fluids with micro needles 21,22 . But none of the existing patches have access to those biomolecules embedded deeply underneath the skin (>1 cm). Importantly, those biomolecules in deep tissues should have a stronger and faster correlation to the dynamic processes inside the human body 23,24 . The wearable photoacoustic patch reported in this work adds an extra sensing dimension for chemical signals in the human body using soft electronics. Furthermore, our wearable patch can map and monitor core temperature in deep tissues with high accuracy and quick response. In existing literatures, core temperature can only be detected by an invasive catheter 25,26 or by heat-flux-model-based wearable temperature sensors 27,28 . However, the latter methods have slow response speed (about several hundreds of seconds) and lack capability of temperature mapping.
In photoacoustic imaging, we developed the low form factor photoacoustic patch. Conventional photoacoustic imaging systems use a laser source to generate ultrasound waves in tissues, and ultrasound transducers to receive photoacoustic waves. Laser equipment is always very bulky and heavy, not suitable for wearing. Safety regulations also require the operation of lasers by professionals. Even though some reported studies introduced the applications of laser diodes [29][30][31][32] or LEDs 33,34 as laser sources, nobody reported using vertical cavity surface emitting laser (VCSEL) bare dies, which have small thickness, ~200 micrometers, and are thus challenging to be integrated on the soft electronic platform. Furthermore, all the reported photoacoustic studies utilize bulky ultrasound transducers to receive photoacoustic waves. Our work introduced a flexible and stretchable ultrasound transducer array as the sensing components. This work develops a flexible and stretchable photoacoustic patch by redesigning and fabricating both the laser source and ultrasound transducer elements. This work makes photoacoustic imaging possible for long-term monitoring even in freely moving subjects.
In summary, we developed a flexible and stretchable photoacoustic patch that can image molecules in deep tissues. This patch can map and monitor the core temperature in deep tissues with high accuracy and quick response. This patch is also innovative in terms of the laser source and ultrasound sensing elements of photoacoustic imaging systems, all in low form factors. None of these advances have been reported by our group or any other groups in the world.

Supplementary Note 3 | Comparison between different photoacoustic imaging systems
For a photoacoustic imaging system, the key sensing components are (1) laser sources for exciting the target molecules to generate photoacoustic waves and (2) piezoelectric transducers for detecting acoustic waves. Conventionally, the optical sources used in the photoacoustic system can be divided into three categories. The first type is the conventional high-power laser system, whose peak pulse power is usually on the order of millijoule 35,36 , with a penetration depth spanning from 3 mm to 4 cm 9 . These high-power lasers are mostly used to provide strong light intensity to excite the target molecules to generate photoacoustic waves. Operating these lasers needs strict training and to be in laboratories that meet high safety standards. Additionally, these laser systems are costly and bulky, which are not suitable for wearable applications. The second type is the handheld compact laser with a relatively lower energy than the first type. A typical laser of this kind has a size as small as 160 mm  64 mm  40 mm 37 . Still, they are too large to be suitable for continuous wearing. The third type is light-emitting diodes or laser diodes. Although some photoacoustic systems employ light-emitting diodes and laser diodes as the optical source, they still rely on bulky rigid ultrasound probes to receive the acoustic waves 29,38 . Those ultrasound probes require manual holding and the subject to be static during testing. Additionally, they use edge-emitting semiconductor laser diodes, which are not suitable to be integrated into a conformal patch because the edge-emitting semiconductor laser diodes usually have a large size (more than several millimeters) in the emission direction 29 .
The photoacoustic patch in this work integrates both the laser source and the piezoelectric transducer into a low form factor conformal patch (20 mm  16 mm  1.2 mm), by encapsulating an array of laser diode chips (1.7 mm  2.4 mm  0.4 mm) and transducer elements (0.6 mm  0.8 mm  1.0 mm) into a flexible and stretchable silicone polymer matrix. In terms of the complexity of the system, we have significantly simplified a conventional photoacoustic imaging system by replacing the bulky laser source with surface-mounted laser diode dies. Although the VCSEL chips (<$10 each) may increase the cost compared to an ultrasound patch, it greatly reduces the cost of conventionally used laser sources. In addition, the cost of each VCSEL die chip can further be reduced if the quantity of chips is increased. The stretchability of the overall patch is enabled by the serpentine shaped metal electrodes that interconnect the laser diode chips and the transducer elements. The device is rigid locally at the laser diode chips and the transducer elements but is soft globally on the system level. The penetration depth of this soft photoacoustic patch in tissues can reach >2 cm. The technology is potentially suitable for wearable health monitoring without immobilizing the test subjects.
To develop a fully integrated wearable system in the future, handling the quantity of data needs to be solved. Different clinical cases require different quantities of continuous imaging for clinical application. The monitoring period extends from several minutes to several days. Assuming each 2D image has a size of 2 cm  2 cm, composed of 200  200 pixels, one 2D image would occupy ~39 KB for 1-byte unsigned integer data type. Thus, 13 slices will be ~507 KB. To be specific, continuous monitoring for 5 minutes, 5 hours, or 5 days will create datasets with the size of about 149 MB, 8.7 GB, and 209 GB, respectively. Such file sizes are easy to be accommodated since common commercial hard disks have space larger than several Terabytes. To handle these data for a completely portable system, a solution is to transfer the image data from the portable system to an external data storage equipment, which can be easily achieved by USB 2.0 cables (data transfer speed > 60 MB s -1 ) or WiFi (data transfer speed > 2.5 MB s -1 ).
It is worth mentioning that the big difference of optical intensity between high power lasers and laser diode chips may have influence on the detection of non-static tissues. Because the high-power laser has very strong light intensity, it can generate strong photoacoustic signals with only one pulse. Pulsation of the tissues (e.g., major arteries) will not affect the imaging result. However, for laser diode chip based photoacoustic systems, the light intensity and therefore photoacoustic signals are relatively weak. Averaging several thousands of signals are required to increase the signal-to-noise ratio. Acquiring several thousands of signals may take one second or even a longer time, during which the photoacoustic signal will move forward and back due to movements of the tissue, resulting in unstable phases of the photoacoustic signals, and thus destroying the coherent averaging (Fig. S46) 39 . Therefore, compensation of motion will be required to achieve a good averaging result for artery imaging 40 .

Supplementary Note 4 | Summary of first, second, and third optical windows
Near-infrared light has high penetration depth in human tissues compared to the visible light because of its weak scattering and absorption 41 . For probing human tissues, three commonly used optical windows are in the range of 650 ~ 950 nm, 1000 ~ 1350 nm, and 1600 ~ 1870 nm 42 . In the first window, hemoglobin still has a higher optical absorption than water and lipid. Therefore, photoacoustic signals of hemoglobin can be generated with low background noise. No extra contrast agent is needed to highlight the hemoglobin. In the second window, the penetration depth increases. But additional contrast agents are needed to label the hemoglobin molecules because of their low absorption coefficients [43][44][45] . The third window has even deeper penetration because of reduced scattering, but is rarely used due to the dominant water absorption 46 , suppressing the detection of other molecules.

Supplementary Note 5 | Comparison between different temperature measurement methods
The gold standard for measuring the core temperature is to use a catheter to measure the temperature in the pulmonary artery 47 , which is too invasive for routine measurements.
Implantable devices with biocompatibility can be directly fixed in the human body, thus providing accurate and continuous temperature measurements in deep tissues 48,49 . However, in a lot of cases, the infection risks, application complexity, data communication, and power supplies of the implantable devices introduce more challenges than benefits.
There are various strategies for noninvasive temperature measurements of the human body.
Wearable skin-like soft sensors usually integrate temperature sensitive electronic components, such as the thermistor 50 , the ion conductor 51 , and the thermocouple 52 . But they can only measure the temperature on the skin surface. Magnetic resonance imaging can quantify the internal temperature variance at a depth >10 cm and spatial resolution of 2 mm 53 . However, owing to the bulky and expensive system, it is not realistic to use MRI in daily activities.
Wearable sensors that can measure core temperatures are developed mostly based on the zeroheat-flux model 54 and the dual-heat-flux model 55,56 . In the zero-heat-flux model, when the skin and deep tissue temperatures are considered identical, there will be no heat flow between them. As a result, the core temperature is the same as the skin surface temperature 57,58 . Nevertheless, these sensors require external heaters to achieve a thermal equilibrium between the skin surface and the core body and thus have a relatively long response time (>180 s) 58 , especially at a considerable depth underneath the skin. To eliminate the use of the heater, sensors based on the dual heat flux model are developed 27 . But this method requires an even longer response time (~447 s) 27 and it is imprecise since it is only a predicted value.
Compared to the existing methods, the photoacoustic patch has multiple advantages, including high penetration depth (>2 cm on tissues), short response time (~1 s), and soft mechanical design for continuous wearing.

Supplementary Note 6 | Mechanism of temperature sensing by the photoacoustic patch
Generating photoacoustic waves is a process of converting optical energy to mechanical vibration energy. After the laser illumination, biomolecules (e.g., hemoglobin in this work) will absorb the optical energy, undergo thermoelastic expansion, and radiate acoustic waves into the surrounding media. For a nanosecond laser source, the generation of photoacoustic waves satisfies the stress and thermal confinements 59 . The photoacoustic signal amplitude can be express as: 60 where Γ is the Grüneisen parameter, µa is the absorption coefficient, and F is the laser fluence.
During the test, the light fluence F is a constant for the same laser source. µa also keeps unchanged for the same type of biomolecule. The Grüneisen parameter is what changes the signal amplitude, and linear to the temperature in the range of 10 ~ 55 °C 59-61 . Therefore, the photoacoustic signal and the temperature show a linear relationship in the vicinity of human core temperature (~37 °C).
The Grüneisen parameter Γ can be expressed as: where Γ0 is the value at temperature T0, α is a constant decided by the tissue type. The photoacoustic signal amplitude can be rewritten as: 62 The photoacoustic signal can quantify the temperature after calibrating αµaF and Γ0µaF, which can be considered as the slope and intercept of a linear function, respectively.
Pure ultrasound techniques can also noninvasively measure the temperature in deep tissue because the tissue temperature will change the sound speed. However, there are some limitations for temperature measurements with ultrasound. First, the biggest problem is that ultrasonography can only detect the contrast of acoustic impedance, which means ultrasound collects anatomical information. As ultrasonography cannot distinguish different biomolecules, it cannot recognize the inclusion components inside cysts, which is critical for determining if the cyst is benign or malignant. Second, ultrasonography may suffer from low contrast to recognize small blood vessels.
Photoacoustic imaging, as a promising biomedical imaging technique, has made a lot of advances in the last two decades 8 . Since the photoacoustic signal originates from the light absorption, photoacoustic imaging holds optical contrast, rather than the acoustic impedance contrast. In addition, photoacoustic imaging combines the best of two worlds: generating signal optically and sensing signal acoustically, which makes photoacoustic imaging best for high-resolution highcontrast imaging of biomolecules in deep tissues. Third, for temperature sensing, ultrasound has a much lower sensitivity than photoacoustics. A quantitative comparison between these two methods has been described. For instance, assuming the temperature of water increases from 20 to 30 °C, the sound speed will increase from ~1481 to ~1507 m s -1 63 , with a relative change of sound speed only ~0.176% per degree centigrade. On the other side, the photoacoustic signal amplitude will be enhanced by 51% for such a 10 °C increase, resulting in a relatively large amplitude change of ~5.1% per degree centigrade 60 .

Supplementary Note 7 | Bland-Altman analysis
Bland-Altman plot analyzes the agreement between a pair of datasets. This plot is widely used in statistics in analytical chemistry as well as biomedicine 64 to compare a new measurement method with the gold standard [65][66][67] . Assuming the datasets measured by the two methods are X and Y, the y-coordinate of the Bland-Altman plot are the differences in each paired X and Y values, while the x-coordinate represents the average value of X and Y. In Bland-Altman plot, there are three horizontal lines, representing the mean bias d ̅ , the upper limit of agreement E upper and the lower limit of agreement E lower . They are defined as follows: where sd is the standard deviation. 1.96 is the boundary of the 95-confidence interval in standard normal distribution. It means that the probability of the population mean value is between -1.96 and 1.96 standard deviations.

Supplementary Note 8 | Characterization of the skin curvature on the imaging performance
To examine the influence of irregular human neck curvature on the imaging performance of the soft photoacoustic patch, the skin curvature distribution was characterized. We used a 3D scanner (HDI Advances, LMI Technologies, Vancouver, Canada) to scan the area above the internal jugular vein (Fig. S44a). The 3D skin surface morphology was reconstructed with high spatial resolution in the software, which was then imported into Catia software (Dassault Systèmes, France) for curvature extraction. Accurate spatial positions of the skin could be read in the Catia.
We acquired 26 typical 1D skin curves by placing 26 planes, with a spacing of 1 mm, vertical to the skin and extracting the intersection line between the planes and the skin surface (Fig. S44b).
Then, the skin curvatures were calculated by circle fitting 68 (Fig. S44c). All of the extracted curvatures are shown in Fig. S44d. The smallest curvature radius is 6.5 cm, which corresponds to the largest deviation from an ultrasound array on a planar surface. The raw irregular skin curves were then used to decide the positions of irregularly distributed ultrasound transducer elements.
To quantify the influence of the skin curvature on the imaging performance, the generation process of the photoacoustic signals was then simulated in a MATLAB toolbox -k-Wave 69 Figure S45a shows the imaging results of the two scenarios side by side. The axial and lateral resolutions for the two scenarios are extracted and displayed in Figs. S45b-S45c. The results show that the average axial and lateral resolutions are only reduced by 0.06 mm and 0.24 mm, respectively, when the transducer array moves from a planar geometry to a curved geometry without phase correction.
Therefore, the irregular skin curvature had a negligible influence on the imaging performance of the soft photoacoustic patch. The reason for this negligible influence is that the working frequency of the photoacoustic patch is ~2 MHz, which is relatively low. The skin curvature radius is not small enough to produce an adverse effect for the long acoustic wavelength in this study.

Supplementary Note 9 | Feasibility of photoacoustic imaging on the detection of different biomolecules
The continuous detection of melanin could have potential applications in close monitoring metastasis of melanoma tumor cells 70 . In addition, melanoma has a very high possibility of metastasis, which causes more than 90% cancer related mortality 71 . Detection and monitoring of metastasis of melanoma tumor cells can help staging the cancer and take effective means of medical intervention at the early stage. Continuous monitoring of circulating melanoma tumor cells has been well studied 70,72,73 . Photothermal therapy has also been used to kill circulating tumors with the assistance of continuous photoacoustic imaging 73 .
For the detection of glucose [74][75][76] , cytochormes 77,78 and nucleic acid, many studies have actually demonstrated both in-vitro and in-vivo label-free imaging using photoacoustic techniques. But for now, photoacoustic imaging is not mature as a reliable technique to continuously monitor humans due to technical and regulatory challenges.
As for exogenous contrast agents, one typical example is indocyanine green (ICG), which has been approved by Food and Drug Administration due to its high biosafety 79 . ICG was not only widely used in photoacoustic imaging studies, but also well established in clinical applications in the field of other optical imaging technique 80 . Specifically for photoacoustic imaging, ICG has been used in the vena mediana cubiti of the right arm of a human volunteer to enhance the monitoring of blood haemodynamics in the finger 81 . In a much more comprehensive study, metastatic status of sentinel lymph nodes in melanoma has been detected with the administration of ICG in 20 patients 82 . The latter study demonstrates that patients can benefit from ICG-assisted photoacoustic imaging for clinical management of melanoma.

Supplementary Note 10 | Comparison of laser pulse repetition rates between different photoacoustic tomography imaging systems
Traditional photoacoustic tomography (PAT) imaging systems usually use expensive high-power lasers as the light source, such as Q-switched Nd:YAG or dye lasers 83 . They usually have a pulse energy ranging from tens of millijoules to several joules and a pulse repetition rate of around 10-100 Hz 9,70,84-90 . One laser shot can deliver sufficient light energy to generate a strong photoacoustic signal with a high signal-to-noise ratio (SNR). Therefore, without the need of data averaging, traditional PAT systems can have an imaging frame rate around 0.5~100 Hz.
However, for small-size low-cost PAT systems, expensive bulky lasers are replaced by compact inexpensive laser diodes or even light emitting diodes (LEDs) 31,38,[91][92][93][94] . These light sources typically have a pulse energy from several microjoules to a few hundreds of microjoules. The laser energy per pulse is not strong enough to generate a measurable photoacoustic signal with an acceptable SNR. Therefore, the data averaging scheme is required to improve the SNR by averaging photoacoustic signals generated by multiple laser shots. Usually, the number of averaging ranges from tens to several thousands, depending on the laser energy. To ensure a high imaging frame rate, a high pulse repetition rate is necessary for the laser diodes or LEDs to achieve the multiple data acquisition, typically in a range from several hundreds of Hz to tens of kHz. In fact, 3 kHz is a typical pulse repetition rate compared with other low-cost PAT systems.
In summary, we need to perform data averaging to improve the SNR because the laser pulse energy is low for laser diodes compared to conventional high-power lasers. To ensure a high imaging frame rate, we need to emit laser pulses at a high pulse repetition rate accordingly. For example, we need to average 3000 times of photoacoustic signals to improve the SNR. Therefore, to ensure an imaging frame rate of about 1 Hz, we have to emit 3000 laser pulses per second.

Supplementary Note 11 | Alternative denoising methods for improving signal-to-noise ratio
In this work, we adopted data averaging and bandpass filter to improve the signal-to-noise ratio (SNR) of photoacoustic signals. The bandpass filter is to eliminate the noise outside the bandwidth of the ultrasound transducers. But thousands of times of averaging makes it time consuming, causing more laser exposure. There are other methods for SNR enhancement while requiring less time consumption, as discussed in the following.
From the hardware perspective, some studies developed code excitation technique [95][96][97][98] to compensate for the low optical energy of laser diodes. Laser diode drivers control the diodes to emit light pulses at a specific sequence. The photoacoustic signal is then acquired by passing through a match filter, which is the so-called decoding. The laser diode drivers usually require a high pulse repetition frequency to satisfy the code excitation. Although the data acquisition time of the data averaging method can also be reduced by using a high pulse repetition frequency, the code excitation scheme consumes even less time 95,97 to achieve the same SNR as data averaging.
From the signal processing perspective, there are also many denoising algorithms to improve the  104 . This adaptive noise cancelling method does not require the prior knowledge of clean signals or noise, compared to conventional adaptive noise filters. All these denoising algorithms can be applied to the photoacoustic signals generated by one laser shot, consuming less time than multiple data acquisition for the data averaging method. But their performance relies on the selection of specific parameters, such as the reserved number of intrinsic mode functions in empirical mode decomposition, threshold in the wavelet thresholding denoising, correlation estimation between the signal and noise in Wiener deconvolution, and time delay for adaptive filtering 104 . The data averaging method and these denoising algorithms can be adopted together to achieve better SNR 99 .

Supplementary Figures
Supplementary Figure 1 | Layout design of the soft photoacoustic patch. The sizes of the laser diodes and the transducers, as well as the spacing between them, are labeled. In data processing, signals of four adjacent elements in the column are summed, digitally connected in parallel to enhance the signal amplitude. Therefore, the 16 rows of transducers form 13 linear arrays during the signal processing. In the column direction, we arrange as many VCSELs as possible to improve the optical energy intensity. In the row direction, we pack as many piezoelectric transducers as possible between VCSELs to improve the image quality. VCSEL: vertical-cavity surface-emitting laser. Unmodified Delay-And-Sum (DAS) algorithm and Coherence-Factor-weighted-Delay-And-Sum (CFDAS) algorithm were applied to reconstruct images, respectively. When Δd equals to 0, the ultrasound transducer array has a uniform pitch between all elements. As the placement of VCSEL chips become sparser, i.e., Δd gets larger, stronger grating lobes are induced. However, the displacement of VCSEL chips also extends the aperture of the linear array, which improves the lateral resolution of photoacoustic images. Therefore, in comparison to the uniformly distributed ultrasound transducer array, VCSEL chips introduce weak grating lobes, but improve the lateral imaging resolution. For the unmodified DAS beamforming algorithm, assuming the photoacoustic signals are measured by a transducer array with M elements, the received signal of each channel is pm(t). To reconstruct the image I(x, z) at pixel (x, z), the wave propagation time from the pixel to the m-th element is ∆tm. Therefore, the image I(x, z) could be computed through the summation of ∑ p m (∆t m ) M m=1

Supplementary
. In the results of applying DAS to beamforming, the grating lobes become larger as Δd increases caused by VCSELs, which degrades images. Therefore, we utilize CFDAS to compensate for this effect. CFDAS introduces an adaptive coherence factor as an additional weight . Both in ultrasound B-mode imaging 105 and photoacoustic imaging 106,107 , the coherence factor weighted DAS beamforming has been demonstrated to suppress the grating lobes. The second row of images show the CFDAS algorithm decreases the impact of grating lobes. c, Images of points targets right underneath the VCSELs reconstructed with CFDAS algorithm. All the images share the same color map, ranging from 0 to 0.8. The images of all targets under the VCSEL have high axial and lateral resolutions, which are only slightly affected by the increased Δd. Furthermore, the intensities of the targets only decrease slightly.

Supplementary Figure 5 | Optical photographs of all the seven layers of Cu electrodes.
The first six layers form the top electrodes, among which layer 6 is for the VCSELs, while others are for the piezoelectric transducers. The bottom electrode is the common ground for all the transducers. Figure 6 | Performance comparison between one element and four parallelly connected elements. a, Schematics of the two measurement strategies. b, The time domain photoacoustic signals received by one element and four elements. c, Comparison of signal amplitudes and signal-to-noise ratios between one element and four elements. Signal-to-noise ratio is defined as 20log10(Peak photoacoustic signal/Root mean squared error of noise) [108][109][110] . Figure 7 | Workflow of the photoacoustic patch. a, Signals received by four elements in the y direction are summed to enhance the signal-to-noise ratio. 13 linear arrays in total are formed equivalently. b, 13 slices of 2D images are reconstructed by the patch. c, A 3D image is formed based on the 13 slices of 2D images. d, After calibration, 3D temperature mapping is achieved, which is the basis for the continuous core temperature monitoring.

Supplementary Figure 8 | Influence of overlap of ultrasound transducers. a, Schematics
showing four cases with the different numbers of overlapped transducers between adjacent linear arrays. b, Imaging planes corresponding to different setups. Three uniformly distributed point sources were set at the depth of 1.5 cm. c, Reconstructed 3D images for different cases. Because of the overlap between adjacent arrays, we can have a large number of linear arrays, which increases the number of 2D images in the y direction. The 3D images show that when there is no overlap, we only have four 2D images. Not all the point sources can be recognized in the 3D image. As the overlap increases, the number of 2D images increases. All points sources are captured. Furthermore, the width of the points in the y direction decreases as the overlapping increases, which means the lateral resolution in the y direction is improved. In summary, increasing the number of overlapped transducers increases the number of 2D images, improving the lateral resolution in the y direction. Figure 9 | Three typical photoacoustic tomography imaging systems. a, Conventional PAT system using bulky ultrasound transducers and expensive bulky high-power lasers 9 . This PAT system has the largest size because the ultrasound probe and the laser are both bulky, immovable, and physically separated. The laser beam is usually guided by optical fibers or lens to illuminate human tissues. The size of the laser is usually over tens of centimeters. The image is adapted from REF 9 , which is licensed under a Creative Commons Attribution 4.0 International License (http://creativecommons.org/licenses/by/4.0/). We added several labels to illustrate the equipment. b, Alternatively, a low-cost photoacoustic probe integrating a traditional bulky ultrasound probe and compact laser diodes 32 . The laser diodes are installed in the ultrasound probe and only need to be connected with an external compact laser diode driver. The integrated photoacoustic probe is handheld and movable. This PAT system reduces the size of the whole system because the laser diodes have much smaller size compared to conventional high-power lasers. The image is copied from REF 32 , which is licensed under the CC BY-NC-ND license (https://creativecommons.org/licenses/by-nc-nd/4.0/). No changes were made. c, Wearable photoacoustic patch integrating low form-factor ultrasound transducers and vertical-cavity surface-emitting laser (VCSEL) diodes. To date, this kind of design has the smallest size in the literature because we greatly reduced the sizes of both ultrasound transducers and laser sources. The laser diodes only need to be connected to an external compact laser diode driver. The patch can be attached to the skin with medical tape, allowing hands-free monitoring. The backing layer decreases the signal amplitude if too thick. We need to make sure the transducers have high sensitivity to receive weak photoacoustic signals, and small thickness not to affect the flexibility and stretchability of the patch. These two factors require the thickness of the backing layer to be as small as possible. On other hand, we still need the backing layer to dampen excessive vibrations to improve the signal bandwidth. In our design, high sensitivity and high flexibility are more important than high bandwidth. To measure the relationships among bandwidth, signal amplitude, and the thickness of the backing layer, we fabricated five different transducer elements, which have the same size of 3 mm  3 mm, but different thicknesses of the backing layer: 0, 0.1, 0.2, 0.6, and 0.8 mm. We excited each element with the same voltage of 50 V and measured the pulse echo signal of each element reflected by an aluminum block. a, Time-domain pulse echo signals of different transducers. b, Frequency-domain pulse echo signals. c, The relationship between the signal amplitude and backing layer thickness. Data are presented as mean values +/standard deviation of 20 measurements. d, The relationship between the -6dB bandwidth and backing layer thickness. Data are presented as mean values +/-standard deviation of 20 measurements. The results show that the amplitude decreases and the bandwidth increases, when the thickness of backing layer increases. In this study, we set the backing thickness as 0.2 mm. Figure 13 | Conformal attaching of the photoacoustic patch to the skin. a, A schematic cross-sectional image of attaching the photoacoustic patch to the skin by using a medical tape. The photoacoustic patch is fabricated by encapsulating ultrasound transducer elements and laser diodes in flexible and stretchable silicone elastomer, i.e., Ecoflex in this work. The ultrasound transducer elements and laser diodes are connected by stretchable serpentine electrodes, in an island-bridge structure. The photoacoustic patch is rigid locally at each ultrasound transducer element and laser diode but soft globally on the system level. Therefore, no external pressure is required to conformally attach the photoacoustic patch to the skin. b, A photo shows attaching the photoacoustic patch on the hand with only a medical tape. A thermocouple (Omega Engineering Inc., SC-TT-K-30-36) was immersed in water to simultaneously measure the water temperature. Signals received by the ultrasound transducer in the patch were recorded at different temperatures from 25 °C to 36 °C, which was the temperature range of the photoacoustic patch after being turned on. The signals' peak-to-peak amplitudes are shown here. The square points are the average of 20 measurements, while the error bars are the stand deviations. The red line is the linear fitting of the average amplitudes. As shown, there is fluctuation at different temperatures. The highest amplitude is about 1.8% higher than the lowest, which can be further reduced by averaging more measurements and transducers. The fitting result shows a 0.8% increase of signal amplitude as the temperature increases. This increase is very small, because the photoacoustic signal typically increases about 5% per degree centigrade 60 . In addition, the temperature of photoacoustic patch reaches a stable state after being turned on for about 8 minutes. We can easily avoid the influence of temperature on the receiving sensitivity if we start the long-term monitoring from 8 minutes after turning on the laser diodes.

Supplementary
Supplementary Figure 17 | Measured receiving sensitivity mappings of the photoacoustic patch at 0 hour and 1 hour, respectively. The 3D image shows the mappings at 21 depths from 5 mm to 25 mm. Two horizontal planes at the depth of 10 mm and 20 mm are also shown on the right, respectively. The receiving sensitivity of the photoacoustic patch was measured in a water tank (ONDA, AIMS III). The results at two moments are very close, which may be because the photoacoustic patch was immersed in water. The heat was dissipated into the water tank. However, according to the analysis in Supplementary Figure 16, the receiving sensitivity of the ultrasound transducers will not change too much either, even when the water temperature rises from 25 °C to 36 °C. Therefore, the dependency of receiving sensitivity on the temperature is not a concern in this study. Figure 18 | Influence of the pulse repetition frequency of the VCSELs on the overall patch temperature. The temperature increases as the pulse repetition frequency increases. Under a given pulse repetition frequency, the patch temperature gradually rises within the first few minutes and then stabilizes. 3 kHz is used in this study. The hardware framework of the system. The Verasonics Vantage 256 controls the timing sequence of the entire system through a customized MATLAB program. It outputs a trigger signal to the signal generator, which then exports a pulsed trigger signal to the laser diode driver. With the driving current, the laser diodes will emit laser and excite the hemoglobin molecules to generate photoacoustic waves. At the moment of laser emission, the transducers start to measure photoacoustic waves and relay the signals to the Verasonics for processing, which is done by customized MATLAB and C programs. b, Timing sequence of the system. The pulse repetition frequency of laser emission and signal receiving is 3 kHz. One frame of image is reconstructed based on 3000 averaged signals in one second. The frequency of measured photoacoustic signal is mainly decided by the laser pulse width and the bandwidth of the piezoelectric transducer.

Supplementary Figure 25 | Influence of external radiofrequency (RF) noise.
We did four groups of tests to show the influence of RF noise, including 1. turning off the power supply and function generator; 2. turning on the power supply and function generator, applying no shielding; 3. turning on the power supply and function generator, and only shielding the power supply and function generator; 4 turning on the power supply and function generator, and only shielding the photoacoustic patch. For the last situation, we assume it is the ideal case since it should eliminate the noise theoretically. We saved and analyzed the background noise at these four conditions. a, 500 frames of time-domain background noise at these four different situations are shown here. Only 500 frames of signals were measured to save storage space. It is obvious that strong RF noise exists in the first three cases, while shielding the patch itself can decrease the RF noise greatly. b, One typical frame of the temporal background noise at different situations. The background noise for the fourth situation is very stationary, only affected by weak RF noise. c, Summation of all of the 500 frames of noise. The summation is also used in the photoacoustic signal measurement, which is equivalent to signal averaging. The summation results show that RF noise is greatly decreased for all cases, benefiting from the data summation. All of the noise curves present the stationary feature, even without any shielding. The RF noise level may be further decreased by more average times. In addition, the noise levels of all of the first three cases are close to each other and only slightly higher than the ideal fourth case. We can also conclude that the power source and function generators introduce neglectable noise, including RF noise and stationary noise.
Supplementary Figure 26 | Influence of the average times on the signal-to-noise ratio. a, Photoacoustic signals averaged by 1000, 3000, and 5000 times, respectively. b, The signal-tonoise ratio increases as the average times. The increased value is calculated by setting the value averaged by 1000 times as the baseline. The frame rate decreases as the average times. 3000 average times are selected in this work because of the trade-off between the signal-to-noise ratio and the frame rate.

Supplementary Figure 27 | Comparison between two different image reconstruction algorithms, Delay And Sum (DAS) and Coherent Factor weighted Delay And Sum (CFDAS).
a, Photoacoustic images are reconstructed based on the DAS and CFDAS, respectively. To characterize the resolutions of a photoacoustic imaging system, a linear object with very small diameter is usually used, such as hairs 91,111,112 , carbon fibers 113,114 . We adopted those wellestablished standard methods for resolution characterization by embedding the linear sources in real biological tissues 112 , water [113][114][115][116][117] , or water like gelatin phantoms 91,111,118,119 . According to the literature, we also find that most studies show the resolution characterization in water or water like gelatin phantoms. Water or water like gelatin phantoms have lower optical absorption and scattering coefficients than realistic biological tissues, which will improve the signal to noise ratio of photoacoustic signals and contrast to noise ratio of photoacoustic images. But they will not affect the characterization results of imaging resolution. Therefore, we used the gelatin phantom as the background media for better accuracy and ease of operation. The hairs were embedded in different gelatin phantoms, respectively. b, Axial and lateral resolutions versus the depth. The CFDAS offers better axial and lateral resolutions than the DAS at all depths and is thus used in this work. To determine the maximum detection depth of the photoacoustic patch, we performed an ex-vivo test on porcine tissue phantoms. We embedded a silicone tube under porcine tissues. The tube was filled with bovine blood. We tested the photoacoustic signals of the tube at the depth of 2 cm, 2.5 cm, and 3 cm. The photoacoustic images were reconstructed and shown. All the images were normalized by the same factor. At the depth of 20 mm, the tube has higher amplitude than the background, which shows good image result. At the depth of 25 mm, the amplitude of the tube is slightly higher than the background. At the depth of 30 mm, the tube and the background media are indistinguishable, which means the photoacoustic patch cannot operate at this detection depth. Therefore, we can determine the maximum penetration depth of the photoacoustic patch is about 25 mm on ex-vivo porcine tissue phantoms.

Supplementary Figure 32 | 3D photoacoustic imaging of overlapping vessels.
To test the capability of 3D imaging of two overlapping vessels, we detected three different gelatin phantoms. a, Schematics of three phantoms, which contain two overlapping transparent silicone tubes, respectively. The tubes are filled with bovine blood. The outer diameters of the tubes are 0.8, 1.5, and 2.0 mm, while the inner diameters are 0.3, 0.5, and 1.0 mm, respectively. Those numbers resemble typical sizes of major vessels in the human body. b, 13 slices of reconstructed 2D photoacoustic images of the overlapping tubes. It is clear that all of the overlapping vessels are distinguishable. c, 3D images of the corresponding overlapping vessels.

Supplementary Figure 33 | Influence of strain on the SNR of the patch.
To test the influence of stretching on the device performance, we tested the device on a gelatin phantom, which contained a transparent silicone tube filled with bovine blood. We quantitatively calculated the signal to noise ratio (SNR) of both temporal signals and photoacoustic images under different strain levels. a, The photoacoustic patch was stretched uniaxially along the horizontal direction by 0%, 5%, 10%, and 15%. Typical temporal photoacoustic signals of one ultrasound transducer are shown when the patch was under different strains. The signals under strains are shifted in time by about 0.5, 1, and 1.5 µs, respectively to make waveforms visually distinguishable. When the patch was stretched, the distance of the laser diodes increased. This could slightly decrease the optical intensity in the phantom. Therefore, the amplitude of photoacoustic signal slightly decreased. SNR of temporal signal is defined as 20log10(Peak photoacoustic signal/Root mean squared error of noise) [108][109][110] . b, SNR of temporal signals has a small decrease as the patch is stretched. Data are presented as mean values +/standard deviation of 10 measurements. However, stretching would also increase the distance between ultrasound transducer elements. As a result, the image reconstruction will be affected if we still use the original transducer positions in the beamforming algorithm. c, The photoacoustic images under different strain levels. As the strain increases, the reconstructed images show stronger distortion. The image quality substantially degraded when the strain level reached 15%. For photoacoustic images, the SNR is defined as 20log10(Average pixel values in the region of object (orange box)/Standard deviation of pixel values in the background (white box)) 86,120 . d, The image SNR will decrease as the strain level increases. Data are presented as mean values +/standard deviation of 10 measurements. In this work, the photoacoustic patch could be attached to the skin with minimal stretching (<5%). The image quality does not show noticeable degradation at 5% strain. For future applications when large strain is required, we can add an additional strain sensor on the photoacoustic patch to monitor the strain level as a way to compensate for the change of distance between transducers. Figure 34 | Relationship between the temperature measured by the thermocouple and the photoacoustic amplitude measured by the patch. To test the ability of photoacoustic patch to measure core temperature, we used the soft photoacoustic patch to measure the temperature in a phantom and checked its performance with thermocouples. a, The phantom is composed of warm bovine blood injected in three regions of interest (ROIs) underneath 2 cm thick room-temperature porcine tissues. Thermocouples were placed in the tubes, where the photoacoustic measurements were also taken for validation. b, Schematic showing the side view of the corresponding phantom. The peripheral regions mean those far from the photoacoustic patch in the horizontal direction, not directly underneath the patch. The porcine tissue phantom is large enough to keep the temperature uniform along the flowing direction inside the tubes enclosed by porcine tissues. c, Beamformed photoacoustic amplitude verse the temperature in the center of ROI 2. Data are presented as mean values +/standard deviation of 20 measurements. Linear fitting (R 2 ~ 0.99) demonstrates the feasibility of core temperature measurement by the patch. Figure 35 | Response of the patch to dynamic temperature changes by continuously measuring warm bovine blood as they were injected into the phantom (Fig. S34). Upon injection, the temperatures of the three regions of interest (ROIs) were close to 32, 36, and 38 °C, which then quickly decreased. The measurement results were validated simultaneously by thermocouples. We injected the same kind of whole bovine blood into the three tubes using three syringes, respectively. Each syringe was filled with warm blood at different temperatures. Therefore, after injecting the blood into the tubes, the initial temperatures in the tubes are different, resulting in different temperature profiles at different moments. a, Photoacoustic images, and b, Temperature mappings captured by the photoacoustic patch at 0, 10, and 25 s after injecting warm blood into the tubes. c-e, Changes in temperature measured by the photoacoustic patch and the thermocouples in ROIs 1-3. The black curves are the mean values of five thermocouple measurements, and the red curves are the mean values of five photoacoustic measurements. The shades of the curves are standard deviations of the measurements. f, Bland-Altman plot (Supplementary Note 7) showing the statistical analysis of 600 pairs of results measured by the thermocouple and photoacoustic patch. The horizontal axis is the mean of the temperatures measured by the two devices, while the vertical axis is the difference between them. Bland-Altman plot analyzes the agreement between two datasets measured by two detection methods 64 , which is widely used to compare a new measurement technique with the gold standard [65][66][67] . Mean defines the average value of difference. −1.96 SD (i.e., standard deviation) and +1.96 SD label the lower and upper limit boundaries of the 95-confidence interval in standard normal distribution. As shown, most (94.8%) of the datapoints are within ±1.96 standard deviations difference, demonstrating the excellent agreement between the two devices with a high statistical robustness. According to our calculation, the standard deviation between the thermocouple and photoacoustic patch is about 0.7 °C, which is considered as the accuracy of the photoacoustic patch at the depth of 2 cm in exvivo porcine tissue. The high accuracy of temperature measurement benefits from the high power of laser diodes and high receiving sensitivity of ultrasound transducers. The entire patch has a pulse energy of about 0.192 mJ with a pulse duration of 200 ns. This power is close to the 0.2 mJ of a commercial LED-array-based photoacoustic imaging system 38,91 , which has achieved in-vivo imaging at a depth of over 2 cm. Comparison of measured receiving sensitivity between the photoacoustic patch and commercial ultrasound probe also shows they have similar performance on the wave receiving (Fig. S22). There are many studies in the literature that have reported accuracies better than this work, such as 0.6 °C in a deep chicken 121 , 0.2 °C and 0.5 °C on phantom and animal using a portable photoacoustic system 122 , and a high accuracy of 0.18 °C in a photoacoustic based close-loop temperature control system 123 . A temperature accuracy of 0.16 °C on porcine tissues has also been reported 60 . Because this is the demonstration of proof-of-concept of flexible and stretchable photoacoustic patch, there is still a lot of space to improve, such as enhancing the laser energy. Figure 36 | Core temperature mapping using the photoacoustic patch. We tested regions of interest (ROI) 2 (Fig. S34), filled with room temperature blood, under a changing thermal gradient created by all of the ROIs. We first injected blood at the room temperature into ROI 2. The blood in ROI 2 was static during the experiment. Then, we quickly injected warm and cold water into ROIs 1 and 3, respectively. The water flow in ROIs 1 and 3 stopped after the warm and cold water filled the tubes fully, which was achieved within 1 second. After the injection, fluids in ROIs 1, 2, and 3 all kept static. We used the photoacoustic patch to image the temperature gradient in ROI 2 created by all the ROIs. a, The mapping results show that the region close to ROI 1 has a much higher temperature than that near ROI 3 initially (top panel); the difference decreases rapidly with time (bottom panel). To verify the accuracy of these results, two thermocouples are placed in the ROI 2, labeled as points I and II. b, The temporal temperature curves measured at these two points show a strong correlation between the thermocouples and the photoacoustic patch. Figure 37 | Schematic showing the experimental setup. We used a pump (Huiyu, BT300J-1A) to drive the blood to flow in a transparent silicone tube with an inner diameter of 3 mm. The rate of flow was set to be ~9 mL s -1 , resulting in a blood flowing speed of ~127 cm s -1 , which was higher than the blood flow velocity of most blood vessels in the human body. The two ends of the tube were immersed in a breaker containing bovine blood. The beaker was placed on a hot plate to heat the blood to different temperatures during flowing. The blood could also naturally cool down while the hot plate was turned off. A portion of the tube was embedded underneath a porcine tissue at a depth of ~2 cm, which was measured by the photoacoustic patch. A soft thermocouple was inserted into the tube to measure the blood temperature simultaneously. The data was read by a thermometer and recorded in a laptop continuously.

Supplementary Figure 38 | Optical distribution simulation at different depths in the water and tissue performed in an open-source MATLAB toolbox -MCmatlab.
For conventional photoacoustic imaging systems, the transducer array is usually immersed in water 9,10 , which serves as the acoustic coupling media between the array and tissue. Ultrasound gel can be also applied as an alternative acoustic coupling media. Water and ultrasound gel both have high optical transparency and low scattering coefficients, which means they have low diffusion capability. In this case, if the optical beam illuminated from an optical fiber or a prism is not uniform, a long stand-off distance (about 10 mm) is usually required to make sure the optical beam reach the tissue skin as uniform as possible. To decrease this stand-off distance or expand the illumination area, an optical diffuser 9,10 can be inserted between the optical source and the tissue. The results in the figure are all normalized individually to show the beam pattern. When the photoacoustic patch illuminates water, the first row presents the distribution of optical intensity at different depths. The absorption coefficient μa, scattering coefficient μs, Henyey-Greenstein scattering anisotropy factor g, and refractive index n are set as 0.00036 cm -1 , 10 cm -1 , 1, and 1.3, respectively. It is obvious that the optical beam is not uniform. This explains why conventional photoacoustic imaging systems need a long stand-off to achieve a uniform beam pattern on the skin surface. The second row shows the optical distribution at different depths in the tissue. The absorption coefficient μa, scattering coefficient μs, Henyey-Greenstein scattering anisotropy factor g, and refractive index n are set as 0.1 cm -1 , 85 cm -1 , 0.9, and 1.3, respectively. The results show that biological tissue has strong diffusion effect on the laser beam. There is a stand-off distance for our patch, which is less than 2.5 mm. The distribution of optical pattern is even more uniform at the depth of 2.5 mm in the tissue than that at the depth of 10 mm in the water. In our studies, we are interested in deep tissues (>5 mm), beyond which the optical distribution is very uniform. Except for the imaging of superficial veins in the hand, foot, thigh, and forearm, all the other experiments are not affected by the stand-off distance. When detecting the superficial veins, a 1 cm-thick gelatin phantom was added between the patch and forearm to compensate for this stand-off distance. Due to the low sensitivity of ultrasound Doppler to slow blood flow, we used a cuff to induce ischemia followed by sudden release of the cuff to increase the blood flow in the vein. Five images were measured before occlusion, during inflation (~ 90 mmHg), right after deflation, after deflation, and no occlusion. Blood flow is undetectable in the first, second, and fifth images. The third image shows the moment immediately after the cuff was released. At that moment, the blood flow was the fastest and thus detected. And the blood flow quickly vanished in the image because the blood velocity dropped fast. The ultrasound Doppler detection of veins in the forearm shows low sensitivity because of the slow blood flow, which is very common for small blood vessels 124 . To the contrary, the photoacoustic patch achieves high-contrast photoacoustic images of the veins. The smallest radius is found to be 6.5 cm, which is used in this work to evaluate the influence of the skin curvature on the imaging performance of the soft photoacoustic patch.

Supplementary Figure 46 | Comparison of averaged signals between consistent and inconsistent phases.
A relatively static tissue yields a consistent phase in the photoacoustic signals, which can get a high signal-to-noise ratio after signal averaging. A dynamic tissue will generate inconsistent phases, resulting in a reduced signal-to-noise ratio after signal averaging.

Supplementary Figure 47 | Influence of the VCSEL chip size on the imaging performance.
Photoacoustic imaging simulation is performed in an open-source MATLAB toolboxk-Wave. a, Schematics showing larger VCSEL chips lead to increased distances between the transducers. Δd is defined as the extra distance caused by VCSEL chips between every three columns of ultrasound transducers. b, Simulated reconstructed images with stronger intensities of grating lobes at increased distances between the transducers. For case Δd = 0 mm, the distance of each two ultrasound transducers (i.e., the pitch) is 0.8 mm, close to the wavelength at 2 MHz. The photoacoustic image shows weak grating lobes. For the other five cases, the effective pitch increases, which enhances grating lobes. However, the amplitudes of grating lobes are still much weaker than the main lobes. Furthermore, the lateral resolution is improved as the effective pitch dimension increases because the aperture of the array is increased. This Figure is correlated to Supplementary Figure 4, emphasizing the influence of VCSEL chip size. Figure 48 | Portable photoacoustic device. The main reason why the current testing system is bulky is that we used a bulky ultrasound research platform (Verasonics system) to acquire photoacoustic signals. Verasonics system also consumes the most power in the entire device. This system is powerful but has many redundant functions that are not necessary for our applications, such as ultrasound wave transmission, high intensity focused ultrasound. The current photoacoustic patch is connected with a high-power laser diode driver and data acquisition system through some cables. Because one coaxial cable for ultrasound signal receiving only has 60 channels, we had to use multiple coaxial cables to connect all of the 240 transducer elements. Alternatively, we can use one coaxial cable containing 256 channels to reduce the number of cables. Therefore, we can keep only two cables connected with the backend controlling system, one for laser diode driving and the other for data acquisition. It is highly possible to replace the current bulky backend system with a handheld backend controlling system. The number of cables will be greatly reduced. The photoacoustic patch, including the laser diodes and ultrasound transducers, can be connected with the portable device with one customized cable. In recent years, portable ultrasound systems with ultra-compact sizes have been developed for point-of-care uses. a, A typical portable system developed by Texas Instruments 125 , which has a size of ~10.0 cm  5.5 cm  1.5 cm. This system can replace the bulky Verasonics system to record photoacoustic signals. In addition, this system has low power consumption of ~2.5 W 125 , which can be powered by a USB cable plugged into a laptop. b, The laser diode driver used in this work has a compact size of ~8.4 cm  4.5 cm  2.0 cm. The power consumption of the driver is less than 10 W when driving the laser diodes at a pulse repetition rate of 3 kHz. Furthermore, the power consumption is tunable and can be further cut down by reducing the laser pulse width, pulse repetition rate, and imaging frame rate. For example, assuming decreasing the imaging frame rate from 1 Hz to 0.5 Hz, the laser diode driver power consumption can be theoretically reduced by ~50%, i.e., <5 W of total power consumption. c, A schematic of handheld photoacoustic device with possible dimensions that are predicted based on existing electrical circuits. The power consumption of such a system can be reduced to be in the range of several Watts.

Wearable
Long In comparison to existing wearable electronics, the photoacoustic patch in this work realizes noninvasive 3D mapping of biomolecules in deep tissues. This technology not only achieves imaging resolutions and detection depth comparable with the conventional bulky systems, but also has compact size and conformal mechanical properties, which are suitable for long-term monitoring.  Table 2 | Comparison between the soft photoacoustic patch in this work and other photoacoustic systems. Specifications of typical photoacoustic tomography systems include laser source type, pulse energy, pulse repetition rate, imaging frame rate, number of data averaging, size of ultrasound probe, and size of laser source. In the literature, PAT systems equipped with high-power lasers usually describe the laser energy per area on the tissue surface, while low-cost PAT systems using laser diodes or LEDs usually describe the pulse energy. The maximum pulse energy for the first kind is in the range from tens of millijoules to several joules by referring to the datasheets of the lasers.

Wearable
Long-term continuous