Evaluation of primary breast cancers using dedicated breast PET and whole-body PET

Metabolic imaging of the primary breast tumor with 18F-fluorodeoxyglucose ([18F]FDG) PET may assist in predicting treatment response in the neoadjuvant chemotherapy (NAC) setting. Dedicated breast PET (dbPET) is a high-resolution imaging modality with demonstrated ability in highlighting intratumoral heterogeneity and identifying small lesions in the breast volume. In this study, we characterized similarities and differences in the uptake of [18F]FDG in dbPET compared to whole-body PET (wbPET) in a cohort of ten patients with biopsy-confirmed, locally advanced breast cancer at the pre-treatment timepoint. Patients received bilateral dbPET and wbPET following administration of 186 MBq and 307 MBq [18F]FDG on separate days, respectively. [18F]FDG uptake measurements and 20 radiomic features based on morphology, tumor intensity, and texture were calculated and compared. There was a fivefold increase in SULpeak for dbPET (median difference (95% CI): 4.0 mL−1 (1.8–6.4 mL−1), p = 0.006). Additionally, spatial heterogeneity features showed statistically significant differences between dbPET and wbPET. The higher [18F]FDG uptake in dbPET highlighted the dynamic range of this breast-specific imaging modality. Combining with the higher spatial resolution, dbPET may be able to detect treatment response in the primary tumor during NAC, and future studies with larger cohorts are warranted.

www.nature.com/scientificreports/ that received both dbPET and wbPET to characterize the similarities and differences in standard descriptive metrics, including standardized uptake values (SUVs), metabolic measures, and spatial heterogeneity statistics.
Since the partial volume effect is driven by the lesion size and spatial resolution of reconstructed images, the cohort was also examined based on the MRI longest diameter (d MRI ) with a cutoff of 2.5 cm [median (IQR)  Table 2). There was a 2.71-fold and 1.58-fold increase in SUL peak in dbPET compared to wbPET in the d MRI ≤ 2.5 cm and d MRI > 2.5 cm groups, respectively. The smaller difference in uptake for the lesions with d MRI > 2.5 cm was expected as the impact on partial volume effect due to the spatial resolution difference between dbPET and wbPET is smaller for larger lesions 14,22 . The differences between dbPET and wbPET in the uptake (SUL) and volume-derived measurements (MTV, TLG) generally increased in the smaller lesions (d MRI ≤ 2.5 cm) than larger tumors (d MRI > 2.5 cm). Finally, 20 tumor morphology, tumor signal intensity, and textural features were generated from both modalities (Table 3). Spatial heterogeneity features, specifically the gray-level co-occurrence matrix (GLCM) correlation

Discussion
The motivation for this study was to compare the performance of primary tumor characterization between [ 18 F]FDG-dbPET and [ 18 F]FDG-wbPET in a pilot study of breast cancer patients. WbPET is primarily used for staging and can detect metastases and nodal involvement, which are necessary to devise treatment strategies. However, due to the collapsed breast volumes and low relative tumor tissue fraction, wbPET may be limited www.nature.com/scientificreports/ in precise and quantitative characterization of the primary tumor for studying therapeutic response. Here, we examine the similarities and differences in the distribution of [ 18 F]FDG uptake in dbPET and wbPET at the pre-treatment time point. Steady-state uptake metrics were used to characterize [ 18 F]FDG-wbPET and -dbPET. As suggested by the PET Response Criteria in Solid Tumors (PERCIST), the lean body mass was used to generate SULs to account for the dual effects of body weight and height on uptake 23 . In general, overall [ 18 F]FDG uptake values (SUL max , SUL mean , and SUL peak ) in the primary tumor were higher in dbPET compared to wbPET. SUL peak is considered more robust than SUL max and SUL mean due to its insensitivity to voxel-wise variations. The higher dbPET SUL peak reflects the sensitivity and dynamic range required for detecting changes in tumor uptake following treatment,  www.nature.com/scientificreports/ although the relative increase in the dynamic range and sensitivity is likely paired with increased background noise. To minimize the effect of background noise, a threshold for dbPET with SUV ≥ 3.0 was applied to define tumor boundaries. The calculation of SUL peak involves mean filtering with a 1.2 cm spherical kernel, which may further reduce the effect of the increased background signal on the presented SUL peak . Although the results did not reach statistical significance, MTV and TLG were estimated to be higher in wbPET compared to dbPET. These trends suggest a role for the partial volume effect, due to lower spatial resolution and tumor tissue fraction in wbPET 14 . Post hoc analysis of SULs and tumor volume supports this analysis, with the observed increases in dbPET-measured SUL peak , MTV, and TLG in lesions with d MRI ≤ 2.5 cm.
To account for variability introduced by the surrounding tissue, the tumor signal is often normalized by the background VOI. In [ 18 F]FDG-wbPET, the background VOI is defined in the liver in the absence of liver metastases, due to its perfused and homogenous uptake of [ 18 F]FDG. However, this normalization method was not available for dbPET and the contralateral breast was used. This method resulted in a fourfold increase of tumor-background ratio in dbPET compared to wbPET using liver as the background. While the contralateral tissue is a logical background tissue for dbPET, careful consideration should be taken if there is bilateral disease. We observed inter-patient heterogeneity in the contralateral breast background signal in both modalities, which resulted in minimal difference in tumor-background ratio following normalization by the contralateral breast in wbPET. Similar variation in background parenchymal FDG uptake have been observed in other studies 24,25 . While limiting the PET scans at the early follicular phase of the menstrual cycle may minimize the background parenchymal signal 26 , the patient's treatment planning should be taken into consideration for scheduling without imposing additional burden. Finally, radiomic features pertaining to heterogeneity and tumor morphology were calculated within the tumor VOI for dbPET and wbPET. The improved spatial resolution within the primary breast tumor and increased tumor tissue fraction in dbPET relative to wbPET have been previously reported, and qualitatively represented in an increase of metabolic heterogeneity in the tumor in a cohort of 35 patients 27 . While textural features of the primary tumor from dbPET images have been analyzed based on breast cancer subtypes 28 , the similarities and differences of these features have not been assessed between dbPET and wbPET. The chosen features have been identified in the literature to be most reflective of changes in spatial and intensity heterogeneity 29 and robust to imaging conditions and reconstruction parameters 30 . To prevent resolution biases from affecting feature calculations, dbPET and wbPET images were down-sampled and up-sampled, respectively, to an isotropic voxel size (2 × 2 × 2 mm). Morphological features were concordant with standard uptake volume metrics (MTV and TLG), especially mesh volume, which was found to be significantly different between the two modalities. Statistically significant difference was also observed in features corresponding to spatial heterogeneity, specifically GLCM normalized inverse difference, joint entropy, and correlation. The differences in textural heterogeneity features correlate to the observed, qualitative differences between dbPET and wbPET and may be primarily driven by the increased sensitivity in dbPET. Table 3. Summary of radiomics features for dbPET and wbPET. Statistical significance, median difference, and 95% confidence intervals were assessed using the Wilcoxon signed-rank test. GLCM gray-level co-occurrence matrix, NGTDM neighborhood gray-tone difference matrix. www.nature.com/scientificreports/ The [ 18 F]FDG-dbPET results in this study are concordant with literature of positron emission mammography (PEM, NaviScan, CA) and previous cohorts with the MAMMI and Elmammo (Shimadzu, Japan) dbPET systems. In a cohort of 388 patients with primary breast lesions, PEM showed comparable accuracy to MRI with improved sensitivity compared to wbPET for quantifying primary lesions 31,32 . Similar to dbPET, PEM had a smaller FOV focusing on the primary lesion at the expense of metastatic lesions, although breast compression and hardware design resulted in limited resolution with maximum intensity projection of 2D images 16 . The sensitivity, spatial resolution, and dose for dbPET in our study recapitulated observations in a cohort of 234 breast cancer patients who received [ 18 F]FDG-dbPET and wbPET/CT for staging 17 . In that study, sub-centimeter lesions in the ipsilateral breast were identified by dbPET that were not resolved in the wbPET images. Similarly, using the Elmammo system, Sasada et al. compared bilateral [ 18 F]FDG-dbPET to [ 18 F]FDG-wbPET in a 47-patient cohort following NAC and observed that the tumor-background ratio was more sensitive than SUV max to treatment response 33 . Nishimatsu et al. also observed higher tumor-background ratios in dbPET relative to wbPET in a cohort of 150 patients, although there was no observable difference in sensitivity to lesion detection 34 . In our study, the SUL peak was higher in dbPET, suggesting a larger dynamic range and potentially, higher sensitivity to treatment-induced changes in [ 18 F]FDG uptake. Furthermore, analysis of radiomic features identified statistically significant differences in GLCM-derived textural heterogeneity markers between dbPET and wbPET.
The limitations of our study are largely driven by hardware design and cohort construction issues. Lesions adjacent to the chest wall experienced limited resolution and incidental scatter from the heart, resulting in inaccurate quantification of uptake. This issue may be mitigated in future studies by using an enlarged aperture on the scan bed, a thinner chest resting area, and a flexible silicone sleeve to gain more breast tissues near the chest wall to be scanned within the detector field of view. An improved reconstruction algorithm as described by O'Connor et al. 18 will be installed to reduce the presence of crosstalk in the breast from FDG signals in the myocardium.
In addition, this pilot study had a limited sample size that prohibited additional analyses based on breast cancer subtypes. While our objective was to compare the [ 18 F]FDG signal distribution in dbPET and wbPET, results from this small cohort cannot be generalized. Study inclusion criteria placed a lower bound on tumor sizes at 2 cm, which limited the study of smaller lesions with high potential for partial volume effect. Since the wbPET was part of a routine clinical procedure, wbPET data included in this report were acquired using different systems with various reconstruction algorithms, depending on clinical availability. SUVs generated across vendors at multiple sites are known to possess coefficients of variation of up to 5%, although this error may be reduced through frequent calibrations 35 . The different wbPET scanners and reconstruction algorithms may also affect radiomics feature stability. To prevent undue adverse effect on reproducibility for wbPET features, we selected more robust and repeatable GLCM features 36,37 in this study. Additionally, tumor-masked images were rescaled to an isotropic resolution and quantized to discrete gray-levels with a fixed bin width 38 to prevent intensity-driven heterogeneity from affecting the radiomics results. Finally, the patient cohort was mostly locally advanced stage II/III patients from a single institution, which may not be representative of the general breast cancer patient population.
In this study, we assessed similarities and differences in [ 18 F]FDG uptake between wbPET and dbPET. SUL peak and SUL max are well recognized as predictors of treatment response and showed consistently higher values measured by dbPET compared to wbPET. While dbPET is not designed to detect metastatic disease for staging, it serves as an adjunct to wbPET and breast MRI for breast tumor characterization 20 . Compared to wbPET, the higher dbPET readout also reflects the higher sensitivity and broader dynamic range for detecting treatmentinduced changes in the primary non-metastatic tumor, providing powerful molecular insights to guide treatment selection and to better assess early molecular changes in response to treatment.
The increased spatial resolution and reduced [ 18 F]FDG dose used in dbPET makes it an attractive modality for treatment monitoring and supports further analyses using higher-order radiomic features to quantify changes in tumor burden and intratumoral heterogeneity for treatment stratification and prediction of survival outcomes. Future studies with this technology would utilize larger cohorts and receptor-specific radiotracers to improve stratification.

Methods
Ethics statement. Ten patients with biopsy-confirmed breast cancer were recruited to participate in an imaging study with [ 18 F]FDG-dbPET (MAMMI, General Equipment and Medical Imaging SA (OncoVision), Valencia, Spain). This study was in compliance with the HIPAA-compliant study protocol that was reviewed by the UCSF Institutional Review Board and approved by the Committee of Human Research under the institution Human Research Program. All procedures performed were in compliance with relevant guidelines and regulations. All patients were required to provide written informed consent to participate.   PET data analysis. To compare wbPET and dbPET [ 18 F]FDG uptake, reconstructed images were first converted to decay-corrected standardized uptake values normalized by body weight (SUV) and lean body mass (SUL). Semi-automated segmentation of the tumor volume of interest (VOI) was performed over the entire volume of the high uptake lesion (OsiriX, Pixmeo, Switzerland) with a threshold of SUV ≥ 2.5 and 3.0 for wbPET and dbPET, respectively. Background SUVs were measured by placing a 1.2 cm cylindrical mask at the centroid of the contralateral breast for dbPET and a 1.2 cm spherical mask in the liver and contralateral breast for wbPET. Following segmentation, the single voxel maximum, average, and peak uptake (SUL max , SUL mean , SUL peak ) were computed as per the standard PET Response Criteria in Solid Tumors (PERCIST), while the metabolic tumor volume (MTV) was calculated as the summing of voxel volumes with SUV ≥ 40% SUV mean 39 . The total lesion glycolysis (TLG) was computed as the product of the MTV and SUV mean .
Radiomic features were computed within the tumor VOIs using Python 3.7. Images were first re-segmented in 3D Slicer 4.11 40 and resampled to an isotropic resolution (2 × 2 × 2 mm 3 ) using linear interpolation and quantized to discrete gray-levels using a fixed bin width 38,41 . Radiomics calculation was performed on these discretized images.
A total of 20 features describing tumor morphology and heterogeneity were calculated using the PyRadiomics package according to the Image Biomarker Standardization Initiative recommendations 42-44 . In particular, heterogeneity features were evaluated from intensity distribution and texture analysis, specifically the greylevel co-occurrence matrix (GLCM) and neighborhood grey-tone difference matrix (NGTDM) 45,46 . Prior to the calculation of textural features, SUV images were harmonized in accordance to literature methods 41 . The data were rescaled to discretized grey levels with a bin width of 0.5. GLCM and NGTDM features were calculated on the 3D volume masked by the tumor VOI with nearest neighbor distances (d = 1) and averaging of each anglespecific matrix. The GLCM and NGTDM classes describe neighboring pixel and regional textures, respectively.
All statistical analyses were performed using R v. 3.6.2 (R Foundation for Statistical Computing, Vienna, Austria). Statistical significance, median difference, and 95% confidence intervals (CI) were assessed using the non-parametric Wilcoxon signed-rank test. p-values less than 0.05 were considered significant. Consent to participate. All patients provided written informed consent to participate.

Data availability
Data will be available upon request.