Highly accelerated knee magnetic resonance imaging using deep neural network (DNN)–based reconstruction: prospective, multi-reader, multi-vendor study

In this prospective, multi-reader, multi-vendor study, we evaluated the performance of a commercially available deep neural network (DNN)–based MR image reconstruction in enabling accelerated 2D fast spin-echo (FSE) knee imaging. Forty-five subjects were prospectively enrolled and randomly divided into three 3T MRIs. Conventional 2D FSE and accelerated 2D FSE sequences were acquired for each subject, and the accelerated FSE images were reconstructed and enhanced with DNN–based reconstruction software (FSE-DNN). Quantitative assessments and diagnostic performances were independently evaluated by three musculoskeletal radiologists. For statistical analyses, paired t-tests, and Pearson’s correlation were used for image quality comparison and inter-reader agreements. Accelerated FSE-DNN reduced scan times by 41.0% on average. FSE-DNN showed better SNR and CNR (p < 0.001). Overall image quality of FSE-DNN was comparable (p > 0.05), and diagnostic performances of FSE-DNN showed comparable lesion detection. Two of cartilage lesions were under-graded or over-graded (n = 2) while there was no significant difference in other image sets (n = 43). Overall inter-reader agreement between FSE-conventional and FSE-DNN showed good agreement (R2 = 0.76; p < 0.001). In conclusion, DNN-based reconstruction can be applied to accelerated knee imaging in multi-vendor MRI scanners, with reduced scan time and comparable image quality. This study suggests the potential for DNN-accelerated knee MRI in clinical practice.


Materials and methods
AIRS Medical provided financial support for this prospective study.The authors had control of the data and the information submitted for publication.

Study population
This prospective study from a single tertiary center was approved by the Institutional Review Board of Yonsei University's Health System (IRB No: 1-2022-0017).Written informed consent was obtained from all enrolled participants.Our study complied with both the Declaration of Helsinki and the Health Insurance Portability and Accountability Act.
Study recruitment commenced from August 2022 to October 2022.Inclusion criteria were: (1) clinically indicated patient for knee MRI; (2) an agreement to participate in DNN accelerated knee MRI; (3) age of 30 years or older; (4) the ability to position the knee in MRI; and (5) symptomatic knees associated with pain

Deep neural network (DNN)-based image reconstruction
Commercially available deep neural network (DNN)-based MR image reconstruction software was used to reconstruct the accelerated acquisition images (SwiftMR, v2.0.1.0.AIRS medical, Seoul, Korea).The software algorithm was based on the popular 2D U-net structure 23 widely used in deep learning architectures in various medical imaging applications.In this model, 18 convolutional blocks, 4 max-pooling layers, 4 up-sampling layers, 4 feature concatenations, and 3 convolutional layers were incorporated in a cascade, with each layer enforcing data consistency.The model was trained and internally validated with 31,865 series and 3540 series of MR images, respectively.The model underwent training using images from the entire body, considering also the musculoskeletal images including the knee.All imaging sequences with different contrasts commonly used in the clinical practice were are included as well.Additionally, the model's loss function was defined as the structural similarity index (SSIM) between the input and the label image, and the model was optimized with Adam 24 over 20 epochs using batch size of 4 at a learning rate of 10 -3 , decaying to 10 -4 .The network was trained using four  The algorithm includes a deep convolutional neural network (CNN) component that removes noise in the image domain and estimates the truncated high-frequency image data.This pipeline can be applied to 2D and 3D acquisitions in multiple anatomic regions and for various pulse sequences, contrast weightings, field strengths, and coil configurations.The amount of noise reduction could be controlled, owing to the model's training process incorporating varying levels of noise amount on the input side.For this study, noise reduction level of low (51% reduction) was used because this level had been found to yield the most similar perceived image quality when compared to the standard images.

Quantitative image quality analysis
For a quantitative comparison of image quality, signal-to-noise ratio (SNR) and contrast-to-noise ratio (CNR) were calculated for all images acquired for this study (conventional, accelerated, and DNN-reconstructed).The femoral bone marrow signal shown in the representative center slice in each image was calculated from a circular region of interest (ROI) placed in the same location for all images, and the mean standard deviation (SD) of the background noise was used for SNR calculation.For CNR, three different imaging plane-landmark combinations were considered-(1) for axial T1-weighted and T2-weighted fat saturated images, marrow-to-muscle signal difference (medial femoral condyle and biceps femoris muscle) and mean SD of background noise were used; (2) for coronal T2-weighted fat saturated images, marrow-to-meniscus signal difference (medial femoral condyle and medial meniscus) and mean SD of background noise were used; (3) for sagittal T2-weighted images, marrow-to-tendon signal difference(distal femur and patellar tendon) and mean SD of background noise were used to calculate the CNR.

Subjective image quality analysis
Three board-certified musculoskeletal radiologists with six years (J.L.), one year (N.L.), and one year (Y.L.) of subspecialty experience individually assessed the conventional and DNN-reconstructed image sets of the 45 knee MRI, using picture archiving and communication system (PACS) monitors (Totoku, Tokyo, Japan).The radiologists were blind to whether the image was FSE-conventional or FSE-DNN.Each image was assigned a unique random number, and the images were evaluated in a random order.The reviewers scored the images independently based on clarity of overall image quality of anatomical structures, perceived image noise, presence of imaging artifacts using a five-point scale.The clear visibility of the medial meniscus and lateral meniscus; ligamentous structures including anterior cruciate ligament (ACL), posterior cruciate ligament (PCL), and medial collateral ligament (MCL); and cartilage grading were estimated separately.

Diagnostic performance for ligamentous, meniscal, and cartilaginous lesions
The three radiologists (J.L., N.L., and Y.L.) formed a consensus on grading meniscal lesions, detecting ligamentous lesions, and grading cartilage by reviewing MR images of non-enrolled patients together before evaluation.The radiologists were blinded to the medical records associated with the images acquired for this study.Three radiologists independently evaluated the knee MRI studies for meniscal, tendinous, ligamentous, and osseocartilaginous injuries.For diagnostic performance analysis, all MRI findings and clinical medical records were reviewed to form a reference standard by two additional reviewers (Y.H.L. and M.J.).
For qualitative analyses of cartilage grading, the radiologists used the Outerbridge classification system 30,31 : cartilage grade, Grade 0 = intact cartilage; grade 1 = signal change on T2-weighted MR images; grade 2 = cartilage defect less than 50 percent of the depth; grade 3 = cartilage defect 50% or more of the depth; and grade 4 = fullthickness cartilage defect with exposure of subchondral bone.When multiple cartilage lesions were present, the cartilage lesion with the highest grade was recorded.

Statistical analysis
Paired t-tests were performed to assess the statistical significance of the difference in the quantitative evaluation of SNR and CNR.For the subjective analysis, we calculated the difference in the qualitative image quality score of anatomical structures, perceived image noise, presence of imaging artifacts by using a paired t-test, and inter-reader agreement was assessed using Pearson's correlation.Diagnostic performances of the FSE-DNN were analyzed in terms of sensitivity, specificity, area under curve (AUC), and accuracy.To assess the diagnostic performance of the images in the cartilaginous lesion, the agreements of FSE-conventional and FSE-DNN were assessed using Pearson's correlation.All statistical analyses were performed in MedCalc (MedCalc Software, Ostend, Belgium) and Microsoft Excel (Microsoft, Redmond, WA, USA).P-values < 0.05 were considered statistically significant.

Subjective image quality analysis
For qualitative evaluation, overall image quality of FSE-DNN was comparable (p > 0.05), depending on the reader.Overall image qualities of anatomical structures were better or equal (n = 42/45 in reviewer 1, n = 44/45 in reviewer 2, and n = 43/45 in reviewer 3).Image noise scores were higher or equal (n = 42/45 in reviewer 1 and 2, n = 44/45 in reviewer 3).Imaging artifacts scores were higher or equal (n = 40/45 in reviewer 1, n = 41/45 in reviewer 2, and n = 42/45 in reviewer 3).The qualitative evaluation of three radiologists were displayed in Fig. 3.The comparisons between FSE-conventional and FSE-DNN images were shown at overall image quality of anatomical structures, perceived image noise, presence of imaging artifacts using a five-point scale.The average and standard deviation of each value are displayed on right side of each score bar.There are no significantly statistical differences (all, p-values > 0.05).Inter-reader agreements of anatomical structures, perceived image noise, presence of imaging artifacts on FSE and FSE-DNN were fair to moderate correlation (R 2 = 0.73, 0.31, and 0.89, respectively; all, p < 0.001).Inter-reader agreement on FSE and FSE-DNN showed good agreement (R 2 = 0.76; p < 0.001).
Representative images for comparable image quality are shown in Fig. 4. The accelerated image exhibited fewer motion-related artifacts, while the TSE-DNN image displayed improved image quality.However, certain artifacts from parallel imaging persisted in the TSE-DNN image, as depicted in Fig. 5.

Diagnostic performance for ligamentous, meniscal, and cartilaginous lesions
All FSE-conv and FSE-DNN images were rated of lesion detection by three interpreting musculoskeletal radiologists.In evaluation of lesion detection, the diagnostic performances of FSE-DNN showed comparable results in ligamentous, meniscal, and cartilaginous lesions (Table 3, Fig. 6).Two of cartilage lesions was under-graded or over-graded (n = 2) while there was no significant difference in other image sets (n = 43).Representative imaging examples for cartilage under-grading or over-grading are shown in Fig. 7.

Discussion
In this post-market multi-vendor study using commercially available DNN-based parallel imaging reconstruction, the FSE-DNN reconstruction of highly accelerated MRI scan reduced acquisition time by an overall 41.0% for a 2D FSE image of the knee MRI.This algorithm is an image-based DNN-reconstruction, which does not require the k-space data nor MRI physics-related information such as multi-channel coil geometry.An older  model of this software has also shown promising generalizability results in pediatric brain 32 and in prostate imaging as well 33 , indicating that an image-based approach may enable accelerated MR exams in clinically used routine sequences with relatively simple modifications such as changing parallel imaging factors and the number of phase encoding lines, etc.This software could restore highly accelerated images in a short time, displaying clinically acceptable image quality and comparable diagnostic performance.This can be differentiated from physics-based (k-space-based) DNN reconstruction 34 which is inherently closely associated with data acquisition methods and may require more computational power and time.As a result, image-based algorithms can be relatively easily deployed in a variety of clinical settings with less MR vendor dependency, and they can even be applied retrospectively to image data in the PACS server.However, further investigation is warranted to accurately compare the image qualities of image-based and k-space-based DNN reconstruction methods.
The strength of our study is that it is a prospective, multi-reader, multi-vendor study as a post-market surveillance.We applied the reconstruction algorithm to knee MRI sequences from all three vendors, showing the possibility of application to multi-vendor MRI applications in radiology.With this strength of this FSE-DNN model, this image-based DNN reconstruction can be easily employed in the radiologic workflow of multi-vendor MRI with various MRI parameters.By changing the parallel imaging factor or number of phase encoding steps of the conventional routine MR sequence, which is easily applicable in the clinical MR imaging protocol, this DNN image reconstruction can reduce scan times with non-inferior image quality and comparable diagnostic performance.
In the quantitative evaluation, FSE-DNN reconstructed images showed higher SNR and CNR, corresponding with previous FSE-DNN studies 35,36 and the same software 32,33 .In subjective qualitative evaluation, FSE-DNN reconstructed images of accelerated FSE images showed non-inferiority against FSE-conventional images in terms of qualitative image quality evaluation.Deep learning reconstruction can be employed in various accelerated imaging techniques [37][38][39] , such as parallel imaging, compressed sensing, or their combination.In our study, we did not compare the combination of compressed sensing and parallel imaging with parallel imaging alone (e.g., CS-SENSE vs. SENSE or CS-SENSE vs. ASSET).Further study on deep learning reconstruction comparison study on combination of CS and parallel imaging is needed in the future.In our study, overall image qualities of anatomical structures were better or equal in most cases.By utilizing deep learning reconstruction of MRI, it is possible to reduce scan time and minimize patient movement, resulting in motion-less imaging.However, the artifacts related with parallel imaging can be pronounced, and these artifacts may persist in some patients.This highlights the need for optimized MRI sequences tailored for accelerated MR imaging.Further research in this direction is imperative in the future.This optimization of MRI sequences involves the choice of CS and the parallel imaging factor, and optimized k-space trajectories.
In the diagnostic performance of lesion analysis, FSE-DNN reconstructed images showed non-inferiority compared to FSE-conventional images.In our study, no significant difference was observed in the diagnostic performance between FSE-conventional and FSE-DNN images.In cartilage evaluation, FSE-DNN showed undergraded lesion (n = 1) and over-graded lesion (n = 1) in small numbers (n = 2/45) among 45 image sets.However, there is no statistical difference between FSE-DNN reconstructed images and FSE-conventional images in this 45-case study.This under-grade or over-grade of cartilage may have originated from acceleration artifacts and image degradation rather than the DNN-reconstruction in our early clinical validation with routine clinical MRI protocols (Fig. 7).Cartilage under-grading on FSE-DNN could also have been affected by the amount of image denoising.This suggests the need for careful selection of acceleration method and denoising settings for cartilage imaging, which may depend on imaging target structures.Conversely, cartilage could be over-graded from parallel imaging-related artifacts.In an under-graded chondromalacia case, cartilage fissuring was smoothened on FSE-DNN images while cartilage signal was slightly enhanced in an over-graded chondromalacia case.This highlights the necessity for MRI sequence optimization, particularly emphasizing the need for more precise learning when it comes to small structures like cartilage and structures influenced by MR signal intensity.This acceleration optimization could be different depending on the target joint (e.g. a large off-center shoulder and a small extremity hand) and target structures such as ligaments, bone marrow, meniscus, and cartilage.Further study involving a larger number of images is needed to validate this aspect.
There were several limitations of this study as well.First, the acquisition parameter modifications were not the same between the three vendors.For example, compressed sensing is routinely used in only one of the scanners, whereas the other two scanners utilize parallel imaging only.We intended the DNN-based reconstruction application to the current MRI sequences, reflecting the clinical practice.Secondly, we set a denoising level of low (51% reduction) according to a preceding internal study on noise reduction level for knee MRI.However, optimal image reduction level should be further investigated in clinical MRI, which could depend on the scanner, imaging joint or target, and radio-frequency coil.Thirdly, our diagnostic evaluations were not confirmed arthroscopically in all patients.We conducted this study with radiologists' consensus as a gold standard of diagnostic performance.Despite evaluation in limited number of patients and limited pathologic findings, this prospective study supports the possibility and the generalizability of DNN reconstruction of highly accelerated MRI to reduce the MRI scan time in clinical practice.Fourthly, the measurements for SNR and CNR were conducted by conventional approach.While optimized methods are available for sensitivity map-based parallel imaging 40,41 , it was necessary to adhere to the conventional approach for practical reasons, as multiple acquisitions were challenging to perform on actual patient images.And, in the context of deep learning-based MRI reconstruction, an optimized method has not yet been established.Further research in this area is warranted in the future.
In conclusion, DNN can be applied to accelerated knee imaging in multi-vendor MRI scanners, with reduced scan time and comparable image quality.Cartilage grading could be under-or over-graded while good agreements in ligamentous and meniscal evaluations.Therefore, the readers should be cautious in utilizing in DNNaccelerated MRI for some lesion evaluation.This study suggests the potential for routine MRI protocols applications to DNN-accelerated knee MRI in clinical practice.

Figure 1 .
Figure 1.Flow chart for prospective study enrollment to evaluate conventional FSE (FSE-conventional) and accelerated MR sequences with DNN reconstruction (FSE-DNN).

Figure 3 .
Figure 3. Qualitative evaluation of three radiologists.The comparison between FSE-conventional and FSE-DNN images was shown at overall image quality of anatomical structures, perceived image noise, presence of imaging artifacts using a five-point scale.The average and standard deviation of each value are displayed on right side of each score bar.There are no significantly statistical differences (all, p-values > 0.05).

Figure 4 .
Figure 4. Conventional and reconstructed images of accelerated sequences: a 47-year-old male with knee pain (A-D).FSE-DNN shows comparable image quality with reduced scan time.The first row represents FSE-conventional images, the second row represents accelerated FSE sequences, and the third row represents DNN-reconstructed images of FSE-DNN.Each column represents the images reconstructed by axial fatsaturated T2-weighted image, sagittal T2-weighted image, coronal fat-saturated T2-weighted image, and axial T1-weighted image.

Figure 5 .
Figure 5. Conventional and reconstructed images of accelerated sequences: a 64-year-old woman with knee pain (A-D).Motion-related artifacts in conventional FSE (A upper, arrow) is not seen on accelerated image (A middle), and the image is enhanced on FSE-DNN image (A lower).Overall image quality is comparable in both conventional FSE and FSE-DNN images (B and D).However, parallel imaging artifacts cannot be completely removed in FSE-DNN image (C middle and lower, arrowheads).The first row represents conventional FSE images, the second row represents accelerated FSE sequences, and the third row represents DNN-reconstructed images of FSE-DNN.Each column represents the images reconstructed by axial fat-saturated T2-weighted image, sagittal T2-weighted image, coronal fat-saturated T2-weighted image, and axial T1-weighted image.

Figure 6 .
Figure 6.Lesion detection and diagnostic performance in conventional FSE and reconstructed images of accelerated sequences (FSE-DNN).(A) A 50-year-old female with knee pain.Mucoid degeneration of ACL (arrows) is shown both conventional FSE (upper A) and FSE-DNN images (lower A). (B) A 60-year-old female with knee pain.Medial meniscal posterior root tear is nicely shown in both images (arrowheads).(C and D) A 58-year-old female and 58-year-old female with knee pain.Cartilage fissuring (arrowheads) and cartilage flaring (boxes) are well delineated in accelerated FSE-DNN images.

Figure 7 .
Figure 7. Cartilage grade on conventional and reconstructed images of accelerated sequences.(A) A 54-yearold female with knee pain.Cartilage fissuring is shown in medial femoral condyle (upper A) while the cartilage fissuring is smoothened, showing under-grade chondromalacia on 2D FSE-DNN image (lower A). (B) A 41-year-old male with knee pain.Cartilage signal changes without significant defect in lateral tibial plateau (upper B) while the cartilage showed T2 high signal intensity defects on 2D FSE-DNN image, showing overgraded cartilage (lower B).

Results Demographic characteristics and scan time reduction
Forty-five patients who underwent this research protocol of knee MRI including routine and accelerated MR pulse sequence were enrolled in three vendors evenly (15 patients for each scanner).The age range of the 45 patients was 30-78 years (mean age ± standard deviation, 53.9 ± 11.8 years).Fourteen patients were male and 31 were female.A total of 45 MRIs of three-vendors were evaluated.Accelerated FSE-DNN reduced scan times by average 41.0% compared to FSE-conventional (GE 41.7%, Philips 43.1%, Siemens 38.1%), respectively.SNR and CNR on accelerated FSE images were significantly decreased with scan time reduction such as changing parallel imaging factors and the number of phase encoding steps.FSE-DNN reconstruction software could enhance the SNR and CNR of accelerated FSE image in a short time.FSE-DNN showed statistically better SNR and CNR than convention FSE : SNR ratio were 2.06, 2.23, 2.63, and 2.10 on axial T2-weighted fat-saturated FSE, axial T1-weighted FSE, sagittal T2-weighted FSE, and coronal T2-weighted fat-saturated FSE, respectively ; CNR radio were 2.14, 3.00, 2.81, and 1.83 of marrow-to-muscle on axial T2-weighted fat-saturated FSE, marrow-tomuscle on axial T1-weighted FSE, marrow-to-meniscus on sagittal T2-weighted FSE, and marrow-to-tendon on coronal T2-weighted fat-saturated FSE, respectively.Scan reductions were 44.63%, 45.56%, 29.80%, and 43.87% in axial T2-weighted fat-saturated FSE, axial T1-weighted FSE, sagittal T2-weighted FSE, and coronal T2-weighted fat-saturated FSE, respectively.The image quality analyses with paired t-test results were summarized in Table2.

Table 3 .
Diagnostic performance of DNN-based reconstruction of FSE-DNN.MM medial meniscus, LM lateral meniscus, ACL anterior cruciate ligament, PCL posterior cruciate ligament, MCL medial collateral cruciate ligament, FCL fibular collateral ligament, BM bone marrow, Sen Sensitivity, Spe Specificity, AUC Area under curve, Acc Accuracy, confidence intervals or p-values in parentheses.