Figure 4. Visualization of segmentation results by different methods and corresponding uncertainty map. Green, red and yellow contour illustrates the ground truth, predictions and overlaid regions respectively.

Figure 3. Quantitative evaluations of different deep learning segmentation models.

Figure 1. Example mediolateral measurements of full-thickness supraspinatus tendon tears made by the algorithm (red) and the reference annotator (green). (a) and (b) demonstrate a case where the algorithm vastly undermeasured the tear due to slice offset and segmentation failure. (c) and (d) demonstrate a different case where the measurements were concordant with similar measurements taken on the same slice.

Figure 3. Histogram of measurement location distances for ML and AP measurements. Measurement location distances are computed (and recorded) as the two distinct distances between corresponding endpoints of algorithm and reference measurements.

Fig. 2 Top Row, sagittal T2-weighted images of the lumbar spine categorized by sequence demonstrated markedly improved image quality using the DLRecon algorithm for 3D imaging: A: 2D T2 standard of care (SOC) T2-weighted fast spin echo (T2w-FSE) B: 3D standard of care (SOC) T2w-FSE C: Deep learning reconstructed (DLREcon) 3D T2w-FSE Bottom Row, axial T2-weighted images of the lumbar spine, categorized by sequence, demonstrate superior image quality with the DLRecon algorithm for 3D imaging: D: 2D T2 standard of care (SOC) E: 3D SOC T2w-FSE F: DLRecon 3D T2w-FSE

Fig. 4 Multiplanar reformations of 3D DLRecon T2 lumbar spine in a patient with moderate scoliosis demonstrating the ability to optimally evaluate each level using orthogonal planes.

Overview of the pipeline. (A) A bone and a cartilage segmentation model ensemble were trained on manually segmented 3D-DESS and used to segment bone and cartilage from 7,437 3D-DESS. (B) Bone shape and cartilage thickness maps were obtained from segmentations. Cartilage T₂ values were parametrically fitted after registering 3D-DESS cartilage to matching MSME. Each biomarker was projected onto the articular bone surface and transformed into spherical coordinates (C). (D) The merged spherical maps were used to train classifier models to diagnose chronic knee pain.

Grad-CAM activation spherical maps overlaid on the articular bone surface for the tibia and patella cartilage thickness and cartilage T₂ models respectively. (A) Tibia Grad-CAM spherical map and projection shows a high activation in the medial tibia. (B) Same tibia cartilage thickness spherical map and projection shows cartilage thinning in the medial tibial cartilage. (C) Patella Grad-CAM spherical map and projection shows a high activation in the medial patella. (D) Same patella cartilage T₂ spherical map and projection shows elevated T₂ values in the medial patellar cartilage.

Table 1. Dice coefficients and surface distances for models with different loss functions and activation functions at different layers. The highest performances are bold.

Figure 1. Prediction masks of two models using categorical cross entropy loss with and without surface distance loss. Adding surface distance loss function significantly improved the segmentation performance.

Figure 1: The design of the self-learning physics informed neural net. Top: a single encoder for each parameter, which consists of 4 fully connected layers with ELU activation layers that end in a sigmoid activation layer to constrain the parameters. Bottom: The full neural net of which the inputs, i.e. the b-matrix and the voxel signal vector, are indicated in orange. The encoders estimate the parameter vector which is then used in the signal generator. The difference between the generated signal and the input signal is minimized by a mean squared loss layer.

Figure 2: Parameter maps obtained from the IVIM-DTI model of one volunteer. The parameter maps look similar but subtle differences can be seen. Especially the signal IVIM perfusion fraction is more homogeneous and higher when obtained with the NN.

Comparison of manual and automatic segmentations from both models and respective 2D unrolled T₂ maps in the right knee of a clinical patient in study 4. Also shown are the average T₂ values from the superficial and deep cartilage regions, cartilage volumes, and DSC scores for the qDESS-trained and OAI-trained models. Arrows indicate examples of visually apparent differences in the automated segmentations and resultant T₂ maps. These differences typically appear at the periphery of tissues, which have limited impact on subregion estimates.

Bland-Altman plots for deep, superficial, and total femoral cartilage T₂ relaxation times for both the OAI-trained and qDESS-trained models. Data is further stratified by study and anterior/central/posterior anatomic region. The T₂ variations are minimal for both models and show no systematic error, however the limits of agreement for the qDESS-trained model for all cartilage layers are smaller.

Entire network structure of the Bloch equation-based autoencoder regularization generative adversarial network (BlochGAN).

Comparison between the reference and the generated T2 fat saturation images from dataset 1. Rows 2, 4, and 6 are the magnified images of the areas in rows 1, 3, and 5 indicated by red arrows.

Figure 2: Grad-CAM map highlights region of interest to classify slice indicating the femoral condyle region. The inverted Grad-CAM map was used to obfuscate the region of interest (femoral condyle) to generate adversarial augmented image. Notice how these variations in intensity could easily mimic for example, RF intensity bias or coil failure. Testing on these augmented images resulted in misclassification in the prediction using the model trained on normal images (Model A), but were correctly classified when trained on adversarial augmented images (Model B).

Figure 3: Other representative examples which were correctly predicted by the model trained on adversarial augmented images (Model B), while being incorrectly predicted by model trained on the original data (Model A).

Figure 2. Fat-suppressed PDw axial images for (a) C-SENSE reference with acceleration factor 2, (b) SENSE (c) C-SENSE, and (d) CS-AI with acceleration factor 4. The yellow arrows denote suprascapular nerve.

Figure 4. Fat-suppressed T2w coronal images for (a) C-SENSE reference with acceleration factor 2, (b) SENSE, (c) C-SENSE, and (d) CS-AI with acceleration factor 4. The yellow arrows denote synovial fluid.

Figure 1. Receiver operating characteristic curves with area under curve (AUC), accuracy, sensitivity, and specificity of the neural network phenotype classifiers. Metrics reported in mean ± standard deviation.

Table 4. Association between phenotypes and longitudinal OA outcomes. We only considered bone and meniscus/cartilage phenotypes in structural OA analyses because the number of baseline knees with inflammatory and hypertrophy phenotypes who acquired structural OA at 48 months were 3 and 2, respectively.

Illustration of the deep learning model. The process was split into two way to solve the weight-imbalance problem and improve efficiency of the model. Modified inception model and UNET was used to detect presence of knee cartilage. In segmentation stage, ‘Modified UNET’, which means custom weight function and additional fully-connected layer applied UNET, was used.

Figure 1. Flowchart of the 3D U-Net model. Numbers on the left side denote the resolution of the tensors, while the numbers on top of the cubes signify the number of features. The left side of the diagram denoted the model’s contracting path. The input-images were the 3D stacks of B1 corrected water and fat images. There were 14 output classes for the thigh images including 11 muscles, plus sciatic nerve, femoral marrow, and background. For the calf images, there were 13 classes including 9 muscles, plus tibial nerve, tibial marrow, fibular marrow, and the background image.

Figure 2. Representative muscle segmentation results. Images were from data in the testing group. The individual muscles were combined to be compartments, and then the whole muscle. Color-coded binary masks of individual muscles, muscle compartments, and the whole muscle are overlaid onto the fat fraction image. The same color codes for each of the muscles are used in the results of the dice coefficient, Bland-Altman, and Pearson correlation analyses.

Figure 4. Segmentation output fused over CUBE

Figure 1. Output.csv and visual representation in the pdf file

Figure 1. Architecture of ResNet50, containing 16 residual blocks. Each residual block begins with one 1x1 convolutional layer, followed by one 3x3 convolutional layer and ends with another 1x1 convolutional layer. The output is then added to the input via a residual connection. The total input number is 6: T1W and T2W of the slice with its two neighboring slices, so one convolutional layer with 1x1 filter is added before ResNet to extract interchannel features and transform from 6 channels to 3 channels as input.

Figure 2. Two true positive malignant cases. The image at left panel shows diffuse tumor infiltration at the 7th cervical (C7) vertebral body with posterior cortical destruction and no apparent collapse. The image at right panel shows diffuse tumor infiltration at third thoracic (T3) vertebra with anterior wedge deformity. The fatty change of other cervical vertebrae in the left panel and T2/T4 vertebrae in right panel is post-radiation effect.

Figure 3. Example images of T1ρ-nosup (a) and T1ρ-sup (c) from conventional curve fitting method, along with the T1ρ map predicted by the deep learning model (b). Synovial fluid (SF) can be easily identified near the knee cartilage on the T1ρ-nosup image (a), while this was largely suppressed on images of the T1ρ-sup (c) and the T1ρ-pred (b). Pairwise absolute difference images (d-f) further illustrate that SF is selectively suppressed (white arrows) without changing the T1ρ of the cartilage.

Figure 2. Workflow for using deep learning to obtain synovial fluid suppressed T1ρ maps from MRI scans without long-T2-selective inversion (LT2SI). Operation A: calculation of T1ρ maps using conventional non-linear exponential curve fitting; Operation B: stack of the T1ρ-nosup image and the TSL source images to form a five-channel dataset as input to the deep learning model.

Figure 1. Network architecture of MSQ-Net and pcMSQ-Net. MSQ-Net with loss_phy to feed back the maps predicted from the model to the input MRI signals, according to equation (6), named physical constraint MSQ-Net (pcMSQ-Net).

Figure 2. Typical results for MSQ-Net and pcMSQ-Net.(a) MRI input signals with different Flip Angles(FAs); (b). T1 maps of GT, and predicted by MSQ-Net, pcMSQ-Net; (c). Difference map from GT. Yellow arrows demonstrates obvious errors could be found in MSQ-Net in some low-signal area while not shown in pcMSQ-Net. (d) and (e), masks of cartilage and meniscus of GT, and predicted by MSQ-Net and pcMSQ-Net.

Figure 2: Comparing images from the standard and fast acquisitions and DL-reconstructed images (Standard, Fast, Fast DL25, Fast DL50, Fast DL75 images). The mean scores from the three radiologists for the ‘overall image quality’ is stated on each image. Large disc protrusion is depicted with the arrow on each sequence.

Figure 1: Comparing images from the standard and fast acquisitions and DL-reconstructed images (Standard, Fast, Fast DL25, Fast DL50, Fast DL75 images). The mean scores from the three radiologists for the ‘overall image quality’ is stated on each image. Changes in multiple vertebral bone marrow can be seen on sagittal images (depicted by solid and dashed arrows, respectively), and facet hypertrophy (depicted by an arrow) on axial images.

Figure 5: Summarizes the evaluation of second stage: facet classification. (a) Visualizes the entire evaluation pipeline where a patch is passed as input (b) Visualization of the model's predictions via saliency maps shows clinically valuable features being highlighted- image above highlights the superior articular portion of the facet as well as the ligamentum flavum; image below highlights the superior and inferior portions of the facet, synovium; (c) ROC curve highlighting AUC, sensitivity and specificity at various operating points along with their confidence intervals.

Figure 3: Summarizes the evaluation of first stage: zero-shot facet detection. (a) visualizes location coordinates annotated on the T2-w axial slices by a neuroradiologist. These location coordinates were used purely for evaluating our localization, and not for training our models; (b) visualizes ground-truth bounding boxes generated from the location coordinates in (a) shown in red against predicted bounding boxes from zero-shot detection, shown in yellow.; (c) characterizes the performance with a mAP-IoU graph.

Fig. 4. The estimated parameter maps for selected cartilage ROIs on the reconstructed -weighted images at TSL = 5 ms for R = 5.2. The reference image and the corresponding parameter maps were obtained from the fully sampled k-space data. The mean values and the standard deviations of the ROI maps are also provided.

Fig. 2. The architectures of the networks used in DEMO. (a) the n-th iteration block in Recon-net. (b) the Mapping-net to generate parametric map.

Figure 3: (a) A ground truth FLASH image. (b) An image constructed from DESS and TSE scans to synthesize the image in panel a. The network looked at single patch sizes of 70×70 pixels to determine if the image was real or generated. (c) An image constructed by using a multi-patch discriminator as shown in Figure 1. The single-patch discriminator in panel b creates new structures (solid arrow) and loses contrast (dashed arrow) compared to the multipatch discriminator in panel c. The undesirable creation of new structure is also shown in the zoomed-in panels (d)-(f).

Figure 1: (a) The presented network uses a discriminator that examines different sized patches of an image to determine if they display real or generated data. The patch sizes are designed so that a 2×2 matrix of one patch size (with an overlap of one pixel) has the same size as the next largest patch. (b) The largest discriminator value in such a 2×2 matrix is selected and compared to the corresponding next largest patch. The larger value from that comparison is stored for that region and the process then repeated for the next patch size. This is applied over the whole image, using a stride of 16.

Figure 2: A summary of the downstream performance of the SSL networks with both pretext tasks (context prediction and restoration). In most data-limited scenarios, SSL pre-training improved results compared to only supervised training on the same data. The impact of SSL was larger for smaller tissues such as the patellar cartilage and meniscus. Smaller patches also provided improved performance.

Figure 1: Examples of image corruptions for context prediction and context restoration for different patch sizes, and the inpainting network’s predictions given the corrupted images as input.

Figure 1. Comparison of the Receiver Operating Characteristic (ROC) curves for varying numbers of training samples, i.e. n=500 (blue), n=1000 (turquoise), n=1500 (green), and n=2493 training samples (purple) in algorithm-based detection of ACL (a) and meniscus tears (b). For the detection of the ACL tears, the area under curve (AUC) increased from 0.64 (n=500) to 0.80 (n=2493). For the detection of meniscus tears, the AUC increased from 0.68 (n=500) to 0.75 (n=2493).

Figure 2. Receiver Operating Characteristic (ROC) curves and corresponding area under curve (AUC) for the test set depicting the difference in performance when training the neural network with expert (purple) vs non-expert (yellow) annotations. Neither in the case of ACL tears (a) nor in the case of meniscus tears (b) does the network benefit from the additional expert annotations by a board-certified radiologist.

Fig. 5. Representative examples of the model inputs (T₁ and T₂ images), radiologist-annotated ground truth segmentations, and the predicted Modic maps. The mapping technique is advantageous for visualizing heterogeneity and transitional pathology.

Fig. 1. Schematic of the full Modic mapping approach. Vertebral bodies are first segmented and extracted from T₁-weighted MRI, allowing extraction of the bodies on the T₁ and registered T₂ images. Next, a binary segmentation network localizes and detects regions of Modic changes. Lastly, each voxel of the detected regions is classified to a Modic type using a nearest neighbor algorithm and T₁ and T₂ z-scores to form a Modic map.

Figure 1: Visualization of segmentation results from each Network.

The first, second and third columns show examples of vertebral body, intervertebral disc, and paraspinal muscle segmentation results, respectively.

Figure 3: Correlation (left column) and agreement (right column) between muscle CSA from manual versus inferred segmentations on each paraspinal muscle.

Agreement is displayed using Bland-Altman plots for CSA on each disc. Correlation between CSA from manual versus inferred muscle segmentations is displayed using a scatter plot, where the line x=y is indicated in grey and each point is the CSA calculated on each respective muscle (both left and right) on each slice in each patient.

Figure 2: The results of automatic segmentation. The red contours are the cartilage regions and the green lines are the subchondral bone regions. (a)the lateral calcaneal surface of the subtalar joint, (b)the lateral talar surface of the subtalar joint, (c)the lateral talar surface of the tibiotalar joint, (d)the lateral tibial surface of the tibiotalar joint. (e)the medial calcaneal surface of the subtalar joint, (f)the medial talar surface of the subtalar joint, (g)the medial talar surface of the tibiotalar joint, (h)the medial tibial surface of tibiotalar joint.

Table 1：Selected features and their corresponding coefficients in the final model. C1-C8 represent eight cartilage ROIs, S1-S8 represent eight subchondral bone 5mm ROIs, W denotes wavelet transform, L denotes Laplacian of Gaussian filtered.

Figure 1 3D ZTE images of the A) ankle without deep learning B) ankle with deep learning, C) hip without deep learning, and D) hip with deep learning.

Figure 3 3D modeling of segmented anatomy showing the A) calcaneus without deep learning, B) calcaneus with deep learning C) 3D printed calcaneus model; and the D) femur without deep learning, E) femur with deep learning, and F) 3D printed femur model. Both 3D printed models were printed using the deep learning reconstruction on a material extrusion printer (Ultimaker S5, Ultimaker, Utrecht, Netherlands).

Figure 1. Overview of the proposed BML lesion identification deep learning model trained with patient-level weak labels.

Scheme of retrospective tuning. From a single T1-weighted image, tissue relaxation parametric maps (T1 map, proton density map, and B1 map) can be predicted using deep neural networks; which are subsequently used to calculate signal intensity of other images (corresponding to different imaging protocols) via the application of Bloch equations.

Retrospective tuning of tissue contrast in MRI. (a) Given a single T1-weighted image acquired at 30°, images presumably acquired at 5°, 10°, and 20° are predicted and compared with the ground truth images, where high image fidelity is achieved in the predicted images. (B) Quantitative evaluation of variable contrast image predictions. Low L1 error (between 0.04 and 0.09) and high correlation coefficients (ranging from 0.97 to 0.99) are consistently achieved.

Figure 2. ROC curves with AUCs for training, validation and test respectively.

Figure 1. Sample femoral cartilage segmented sagittal fat-saturated proton density weight clinical MR images.

Figure 1 A 24-year-old male suffered multiple injuries caused by traffic accident for more than 3 days. Bone contusions (red arrow) showed striped low signals on T1 sequence (Figure a), while striped high signals on PD-TSE (Figure b-c), DWI (Figure d) and ADC (Figure e). The injury of the medial meniscus (yellow arrow) displayed small patchy high signals on PD-TSE (Figure f). *ADC value of ROI: 794.6 (× 10^-6 mm²/S)

Figure 2 A 62-year-old female suffered from pain of the right knee with no obvious causes, accompanied by limited movement for more than 3 years. The bone marrow lesions of OA (red arrow) showed small patchy and irregular low signals on T1 (Figure a), while small patchy high signals on PD-TSE (Figure b-c), DWI (Figure d) and ADC (Figure e). Cystic degeneration of bone marrow lesions (blue arrow) displayed round-like low signals on T1 (Figure a) while round-like high signals on PD-TSE (Figure b-c). *ADC value of ROI: 1,278 (× 10^-6 mm²/S)

Figure 1

Figure 2

Representative images for sparse reconstruction of retroactively undersampled MR data. At lower sampling rates, less of the bone microstructure is apparent and more artifacts occlude the reconstructed image.

(A) Binomial probability mass function used to randomly select indices in k-space for undersampling. (B) Mask of sample indices in k-space; white indicates a measurement, black an unused k-space value. (C) Fully sampled 2-D slice of k-space of a typical MR image. (D) Undersampled k-space based on mask indices.

Figure 1: Reconstructed fat, water and PDFF maps using IDEAL and mask used for radiomic computation.

Figure 2: Correlation of first order radiomic features with clinical features and T-score (spine, hip, femoral neck) calculated on fat, water and PDFF maps. IQR: Interquartile range, MAD: Mean absolute deviation, RMS: Root Mean Squared.

Figure 2: A representative UTE-Cones-MT MRI image of the lower leg of a 37-year-old female subject. A representative region of interest is depicted on cortical bone in tibial midshaft (yellow).

Figure 4: Boxplots of (A) MMF, (B) TWPD, (C) BWPD, (D) PWPD, (E) MMPD, and (F) T1 values in OPe, OPo, and Ctrl cohorts. Average, median, SD, first, and third quartiles values are indicated in the boxplots.

Performance of classification with different machine learning methods. Heat map shows the AUC of classification with four classifiers in different imaging sequences. The heat map (located to the right of the entire image) illustrates that the darker the color, the higher the AUC. DT, decision tree; LR, logistic regression; RF, random forest; SVM, support vector machine; FS-T2, fat suppression-T2WI; T1, T1WI.

Performance of radiomics model based on different MRI sequences using LR classifier. Based on T1W sequence, FS-T2W sequence and combined sequences. AUC, area under curve; FS-T2WI, fat suppression-T2WI; T1, T1WI.

Figure 2. Bar graph of adjusted mean differences (age, gender, BMI) in WORMS scores between baseline and 4 years of medial meniscus body for cases of sustained synovitis based on definition (i) and controls. The error bars represent 95% confidence intervals.

Figure 3. Bar graph demonstrating adjusted mean differences (age, gender, BMI) in WORMS scores between baseline and 4 years of bone marrow edema in the lateral tibia for cases of sustained synovitis based on definition (i) and controls. The error bars represent 95% confidence intervals.

Figure 3. AcidoCEST FISP MRI sequence of mouse's lumbosacral region and correspondent immunofluorescence GLUT1 staining. (A) Average pH measurements of L6 (7.0) and S1 (6.8) were obtained with AcidoCEST. (B, C, D) Correspondent immunofluorescence GLUT1 staining shows higher average GLUT1 staining in S1 (2.02x) compared with L6 (1.36x), inversely correlated with the pH measurements.

Figure 4. AcidoCEST FISP MRI sequence of mouse's coccygeal region and correspondent immunofluorescence GLUT1 staining. (A) Average pH measurements of C1 (6.9) and C2 (6.6) were obtained with AcidoCEST. (B) Immunofluorescence GLUT1 staining shows higher average GLUT1 staining in C2 (1.99x) compared with C1 (1.24x), inversely correlated with the pH measurements.

Figure 1: A 25-year-old man with a 2-year history of axSpA. (A) Oblique coronal T1 FSE; (B) and (C) oblique coronal ZTE without and with postprocessing; (D) Oblique coronal CT. ZTE exhibits similar bone contrast to CT and provides a similar depiction of erosions in various areas on CT. ZTE with postprocessing has higher contrast and sharper and clearer edges. The erosions in the inferior quadrant of right iliac are not well depicted on the T1 FSE.

Table 2: Sensitivity, specificity, accuracy and the consistency for erosion detection of three images, compared to CT.

Fig.3. (a)Comparison of average PDFF values of lumbar vertebral among three groups. (b)Relationship of PDFF value and BMD. As the PDFF values increase, the bone mineral density decreases.

Fig.2. Bland-Altman difference plots for PDFF measurements generated by using 1.5T MR and 3.0T MR IDEAL-IQ.

Images in a 38-year-old woman with axSpA.Conventional and synthetic MR images of the sacroiliac joint show the clear presence of BME in the left sacrum articular surface(yellow arrow)that is hyperintense on STIR.The lesion on T1-mapping and T2-mapping show higher value than the surrounding.Conventional and synthetic MR images of the sacroiliac joint show the fat metaplasia in the left iliac articular surface(white arrow)that is hyperintense on T1-FSE.The lesion on T2-mapping and PD-mapping show higher value than the surrounding while T1-mapping shows lower value.

Figure 1: Child mummy in a custom-built Tx/Rx Birdcage coil with fast switching circuitry connected to a clinical 3T MRI.

Figure 2: Direct comparison of the child mummy head images between a) CT image (100 kV) and b) MR image (TE=70 µs). Signal intensities were evaluated in the marked ROIs.

Figure 2. The correlation between BAP and Osteophyte WORMS scores was illustrated by correlation fitting scatter plots. Partial correlations are listed with adjustments for age, gender, BMI, KL grade.

Figure 1 a and b: A 51-year-old famale with OA, right knee K-L 3, PD-weighted image(a) and T2 Mapping image(b), Meniscus lateral anterior is normal and the values of T2 is 24.72 ms, osteocalcin: 19.66 ng/ml, tP1NP: 64.94 ng/ml, β-CTX: 691 pg/ml. c and d: A 71-year-old famale with OA, right knee K-L 2, PD-weighted image(c) and T2 Mapping image(d), Meniscus lateral anterior is degeneration and the values of T2 is 28.44 ms, osteocalcin: 13.51 ng/ml, tP1NP: 46.25 ng/ml, β-CTX: 224 pg/ml.

Figure 3. Left: In Scheme 1, the logistic regression (LR) model has the best performance. Receiver-operating characteristic analysis revealed the superior performance of this model for T1-weighted images (AUC, 0.92). Right: The LR model was not predictive for STIR or concatenated images (AUC, 0.56 and 0.71, respectively).

SVM, support vector machine; MLP, multilayer perceptron.

Figure 5. Receiver-operating characteristic curves for differentiating myelodysplastic syndrome from aplastic anemia using T1-weighted images with Scheme 2. Institutional differences in performance are apparent (area under the curve: University, 0.879; Nishi, 0.962; Chugoku, 0.742).

The typical case of synovitis. MRI of a 29-year-old woman with RA. (a) T2W-SPAIR shows high signal intensity within intercarpal-carpometacarpal joints of the left hand (The region of interest, ROI). (b) Contrast-enhanced MRI shows high signal intensity (score 3: severe enhancements). (c) DKI (b value, 500 s/mm²) shows a hyperintense signal covered (ROI). The ADC map (d), D map (e), and K map (f) show ADC, D, and K values for the lesion (ROI) of 1.528×10^-3 mm²/s, 1.913×10^-3 mm²/s, and 0.603.

The typical case of joint effusion. MRI of a 56-year-old woman with suspected RA. (a) T2W-SPAIR shows high signal intensity within distal radioulnar joints of the right hand (ROI). (b) Contrast-enhanced MRI shows low signal intensity (score 0: no enhancements). (c) DKI (b value, 500 s/mm2) shows a hyperintense signal covered (ROI). The ADC map (d), D map (e), and K map (f) show ADC, D, and K values for the lesion (ROI) of 2.225×10^-3 mm²/s, 2.463×10^-3 mm²/s, and 0.313.

Table 2: Spearman correlation coefficients between regional SUV_max and corresponding regional synovitis severity grades on CE-MRI, overall synovitis grade, MOAKS effusion and Hoffa synovitis grades, and median K^trans values

Figure 1: Side-by-side comparison of sagittal CE-MRI, K^trans, and SUV_max fusion images. White arrowheads indicate synovitis at the suprapatellar pouch, while clear arrowheads indicate synovitis at the posterior surface of Hoffa’s fat pad.

Figure 1. 3D-HR-PD imaging of the knee joint for a 25 year old healthy volunteer: a. CS0; b. CS4; c. CS6; d. CS8; e. CS10; f. CS12. Visualization of anterior cruciate ligament (ACL) was significantly worse at CS=10 and 12.

Figure 2. a.CS 0 3D 3D VIEW PD. b. CS 4 3D 3D VIEW PD. c. CS 6 3D 3D VIEW PD. d. CS 8 3D 3D VIEW PD. e. CS 10 3D 3D VIEW PD. f. CS 0 3D 3D VIEW PD 25 year old healthy volunteer, the posterior cruciate ligament was significantly worse at CS=10 and 12.

Figure 1. (A) Coronal MAVRIC inversion recovery (IR), (B) T₂ mapping, and (C) apparent diffusion coefficient (ADC) images from a 67 y.o. female subject with a synovial reaction classified as ‘metallosis’ synovial impression. Arrows point to regions of interest selected for ADC and T₂ mapping analysis.

Figure 2. ADC and T₂ values for ‘Abnormal’ and ‘Normal’ classifications.

Figure 3. Abnormal findings on MAVRIC-SL MRI to suggest periprosthetic joint infection

Figure 5. 70s, female. Periprosthetic joint infection of right hip arthroplasty. Therapeutic surgical intervention was needed. (A) 2D FSE T2WI transverse image, (B) 2D FSE STIR transverse image, (C) MAVRIC-SL STIR coronal image. There is a large signal loss at the joint area on 2D FSE T2WI and STIR. Although periprosthetic assessment is not available on 2D FSE T2WI and STIR, soft-tissue fluid collection is noted on 2D FSE T2WI and STIR (white arrows). While, soft-tissue fluid collection with communication with the joint is clearly noted on MAVRIC-SL STIR (white arrows and arrow heads).

Figure. A-C: A 62 years old female patient with OA, K/L score=1, VAS=15, WOMAC= 60. Representative sagital images T2FS (A), IVIM (B), and pseudo-color map of the IVIM (C). The ROI was drawn in the IPFP without T2FS-hyperintense regions in IVIM map. D-F: A 66 years old female patient with OA, K/L score =2, VAS= 22, WOMAC=79. Red arrow indicated T2FS-hyperintense region within IPFP on T2 map(D) and pseudo-color map of the IVIM (F), and high perfusion were depicted in red (F). The ROI was drawn in the T2FS-hyperintense region in IVIM map (E).