Figure 1: Architecture of CycleSeg, which is a combination of CycleGAN and 2 Unets. The notations in this figure correspond to those in Equation (1) in the text.

Figure 4: Qualitative comparison of organ segmentation for a representative slice from the four networks (detail in text). The segmentation is displayed as contour.

Figure 2: A case example from an 83-year-old man with prostate cancer (tPSA=7.13 ng/ml, Gleason Score=4+5). The lesion is manually outlined. (A) The first DCE time frame (pre-contrast image); (B) The 15th DCE time frame (post-contrast image); (C) The 40th DCE time frame (post-contrast image); (D) the DCE time intensity curve shows the wash-out kinetic pattern.

Figure 4: Different ROIs from the case shown on Figure 2, which contains different amounts of peritumoral tissue as the input for the diagnostic neural network. (A) Original image, (B) Tumor alone (R0), (C) 120% enlarged tumor area (R2), (D) 150% enlarged tumor area (R3), (E) 5-pixel expansion from the tumor boundary (R4), (F) 10-pixel expansion from the tumor boundary (R5), (G) region growing with ±20% average intensity as the stopping criteria (R7), (F) region growing with ±30% average intensity as the stopping criteria (R8).

Figure 1 The workflow of the research. We trained the PAGNet on training cohort from one-site, and evaluated its performance with cohorts from both institutions. The performance of the proposed model was compared with those of a ResNeXt model and two radiologists.

Figure 4 ROC curves (top) and Grad-CAM of PAGNet (bottom). PAGNet performed better than ResNeXt model and clinical reports in the internal test cohort (a), and achieved results comparable to the radiologists on the external cohort (b). The Grad-CAMs of the model on one positive case (c) and one negative case (d) were also shown, in which the red regions indicate where the model paid attention.

Figure 1: The proposed end-to-end pipeline includes automated intensity normalization using AutoRef⁵, automated prostate segmentation using VNet⁶ and nnU-Net⁷ followed by an automated Quality Control step, and the sampling of cropped images with different techniques and settings. The cropped images were then used to train weakly-supervised (Fixed-Point GAN) and unsupervised GAN (f-AnoGAN) models.

Figure 3: f-AnoGAN and Fixed-Point GAN - 3.a) f-AnoGAN: Linear latent space interpolation for random endpoints of trained f-AnoGAN shows that the model does not focus only on one part of the training dataset. 3.b) f-AnoGAN: Mapping from image space (query) back to GAN's latent space should yield resembled images. Here, the mapped images are similar but not entirely identical. 3.c) f-AnoGAN: Positive test case that fails– 3.d) Fixed-Point GAN: negative and positive tested cases, no differences for the negative case, whereas positive case found, and localized (difference).

Figure 1. The overview of the ResNet-UDM. The output of ResNet50 was continuous and would be discretized with three self-learnt . The discrete output was than compared with the ground truth. In the Inference stage, were also used to discretize the output of ResNet50 and produce the final PI-RADS category.

Table 1. The comparison of ResNet-UDM and S. Thomas’s work.

Segmentation of PG and GTV for test patient 1 - 3 (a) Input mpMRI sequences, (b) ground truth, (c) prediction. Overlays in (b) and (c) depict PG in blue and GTV in red.(d-e) overlay of the saliency map for the class PG and GTV on the corresponding input sequences and (f) the generated saliency map PG and GTV. The locations of higher intensity values in the saliency map indicate the importance of the corresponding voxels in the prediction of CNN.

DSC for Test cohort (n = 15). The red lines in the plot show the median Dice value for the classes BG, GTV and PG. The upper and lower bounds of the blue box indicate the 25th and 75th percentiles, respectively.

Figure 1. The image-to-image translation network translates DW-MRI data from mp-MRI acquisitions to VERDICT-MRI. GAN-type losses are used to generate realistic VERDICT-MRI data from DW-MRI form mp-MRI acquisitions while standard cycle-consistency losses allow us to preserve the content during the translation. A segmentation loss on the cycle-reconstructed images obtained using a pre-trained source-domain network, ensures that critical structures corresponding to the prostate lesions are successfully preserved.

We translate the DW-MRI from mp-MRI acquisitions to VERDICT-MRI and supervise using the ground-truth segmentation masks.

Figure 3: (a) Evaluation of subsampled TE-b schemes for ex vivo prostate DR-CSI. For a fixed total encoding number, each TE-b scheme was compared to the reference 4x7 scheme in terms of the mean total spectrum error (TSpE) in 9 prostates. (b) Selected schemes with the minimum mean TSpE (c) Slice-averaged T2-D spectra and signal component fraction maps from each schemes. The definitions of the three spectral peaks are indicated on the reference T2-D spectrum. The prostate cancer region of interest (transition zone) is denoted by the black contour on the reference (f_A,, f_B, f_C) maps.

Figure 5: (a) Evaluation of subsampled TE-b schemes for in vivo prostate DR-CSI. For a fixed total encoding number, each TE-b scheme was compared to the reference 5x4 scheme in terms of the mean total spectrum error (TSpE) . (b) Slice-averaged T2-D spectra and signal component fraction maps from each selected schemes. The definitions of the four spectral peaks are indicated on the reference T2-D spectrum (c) Evaluation of signal component fraction estimates in the overall prostate peripheral zone (PZ) and transition zone (TZ). compared to reference

Figure 1: Shown here is the cine acquisition in the sagittal plane. It can be seen that as the subject breathes, the prostate also moves by a small amount with the breathing motion, especially along the feet/head direction.

Figure 3: Shown here are the power spectra for the axial scan with diffusion in the S/I direction. The linear phase is being measured in the A/P and R/L directions. The constant phase, linear phase in the A/P direction and the residual phase after subtracting the linear and constant phases show frequency peaks at similar frequencies to the breathing curve. Note that the values around 0 of each power spectrum were removed prior to calculating the normalized power, which was found by taking the square root of the sum of squares of the power spectrum.

Figure 1. (A) Mean DW Image (900 sec/mm²) of patient 6. (B) Standard deviation per voxel image for patient 6. (C) Mean signal intensity versus acquisition # divided into three diffusion-sensitizing gradient directions (Black line = mean, Dashed error bar = margin of error, Dashed gray line = range). White arrow points to cancerous lesion while red arrow points to noise.

Table 1. Mean Signal, % Range, and SNR of patients 1-7 for a 3x3 region of voxels in cancerous and benign contralateral tissue.

Figure 1: (A): A diagram of the multi-readout DWI sequence. Multiple EPI echo-trains (or readout trains) are incorporated into the sequence, with each echo-train corresponding to a specific effective TE. (B): In the multi-readout DWI sequence, signal excitation is accomplished by a 2D RF pulse to restrict the FOV so that a shorter echo train length can be used in each echo-train, allowing multiple echo-trains to be acquired.

Figure 4: ADC maps of the prostate at different TEs overlaid onto the corresponding T₂*-weighted images (b-value of 0 s/mm²) of a healthy subject (top row, average values: 1.49, 1.58, and 1.67 μm²/ms) and a patient with prostate cancer (bottom row, average values: 1.35, 1.39, and 1.45 μm²/ms). The red arrow indicates the focal region of the cancer.

Figure 2. Prostate cancer in transitional zone. Cancer lesion is shown as a homogeneous hypointense lesion with mass effect on T2WI (arrow). synDWI₁₀₀₀ with shorter TR provided better contrast than DWI₁₀₀₀ with long TR, whereas noise in synDWI₁₀₀₀ with TR of 500ms become more visible.

Figure 4. CR (a) and visual score (b) of prostate cancer between synDWI₁₀₀₀ with shorter TR of 1000ms and DWI₁₀₀₀ with long TR of 6000ms in three patients.

Fig 3. Images from three representative subjects.

Fig 2. Alternative fits to the same voxel. Note the fitted signals (top row) are similar but the T2 distributions (lower row) differ. LWF maps inset.

Prostate cancer lesion is shown as a homogeneous hypointense lesion with mass effect on T2-weighted imaging (arrow) (A) and focal early enhancement on dynamic contrast-enhanced MR imaging (arrow) (B). Signal intensity of normal prostate is lower in b3000 (D) than in b2000 (C). However, the lesion conspicuity between b2000 (arrow) and b3000 (arrow) is almost equivalent.

Figure 2. Receiving operator characteristic (ROC) curves for conventional ADC (dark gray), RSI-C₁ (green), and PI-RADS v2.x (light gray), for the patient-level detection of higher-grade prostate cancer.

Figure 1. Histograms of (A) conventional ADC (lowest voxel value in prostate), (B) RSI-C₁ (highest voxel value in prostate), and (C) highest PI-RADS category (v2 prior to 2019, v2.1 in 2019). Blue: Patients with no cancer or low-grade cancer. Orange: Patients with higher-grade (grade group ≥2) prostate cancer. Brown: where blue and orange overlap.

Details of the imaging parameters

A 80-year-old male with prostate cancer (PSA level of 13.69 ng/mL, Gleason score of 3+4) in the transitional zone. Cancer lesion is shown as a homogeneous hypointense lesion with mass effect on T2-weighted imaging (A). The lesion with focal hyperintensity is depicted clearly in the three DWI image (sshEPI DWI, sshEPI DWI using C-SENSE, and IRIS) (B, C and D). SNR and CNR is higher in sshEPI DWI than in sshEPI DWI using C-SENSE and IRIS, whereas sharpness in sshEPI DWI using C-SENSE and IRIS is better than sshEPI DWI.

Details of the imaging parameters

A 82-year-old male with prostate cancer (PSA level of 6.56 ng/mL, Gleason score of 4+3) in the right peripheral zone. Cancer lesion is shown as a homogeneous hypointense lesion with mass effect on T2-weighted imaging (A). The lesion with focal hyperintensity is depicted clearly in the both DWI image (ssEPIS and EPICS) (B, C). Signal to noise ratio is higher in EPICS than in EPIS.

Figure2 For BPH (yellow), the hyperintense region was observed on (b) ADC, (c) D, and (e) f; whereas hypointense region was observed on, (a) High b-value map, (d) D*, and (f) k of a representative patient diagnosed with PCa and BPH.

Figure4 Diagnostic performance of IVIM-DKI parameters obtained from HY and HY+TV model to differentiate from (a) Tumor vs BPH and (b) Tumor vs Healthy PZ.

Figure 1. The diffusion-weighted images of prostate by ssEPI-DWI and MUSE-DWI.

Figure 2. The data analysis procedure. (a) The prostate ROI was segmented manually from high-resolution structural images, and (b) the overlapping between the segmented prostate ROI and prostate range by DWI images was calculated.

Fig.1. A dot plot of AUC values of receiver operating characteristics in differentiation of presence or absence of clinically significant prostate cancer for 2D-small-ROI with different area

Fig.5. Bland-Altman plots of ADC values measured by using 3D-whole-lesion-ROI method (10th percentile). 95% limits of agreement among the readers on was +/-112.

Figure 2: The T2-weighted image, ADC map, shortest-TE HiSS MRI image, and HiSS MRI-derived temporal domain (R, R1, R2) and spectral domain (PW, PRD, PRA) parameter maps are shown here for a representative slice through the prostate of a 52-yo man with a Gleason score 7 (3+4) lesion (white outline).

Figure 3: Boxplots of HiSS MRI-derived parameters R, R1, R2, PW, PRD, and PRA are shown, for normal and cancer ROIs. The whiskers represent the full data range. All HiSS-derived parameters were statistically significantly different (p < 0.05) between cancer and normal tissue ROIs, except R2.

Figure 1. Prostate axial ADC (a), ρ (b), V (c), and k_io (d) maps; inferior perspective. The GS (3+4) lesion is bordered in yellow (a). [Pixels in urethra and ductule-rich regions are colored black in (b), (c), and (d).] The ρ, V, maps are in near absolute quantitative agreement with estimations from published ex vivo histopathology and the r map shows very high in vivo lesion/NA prostate gland conspicuity. The k_io map indicates lower metabolic activity in the lesion. The r and V maps may suggest nascent pathology in NA prostate regions, particularly the peripheral zone (see text).

ROI can be obtained without layer-by-layer annotation through the interpolation and annotation tools of AS

De-artifacting workflow

Table 1. Average PIDS prevalence across the 40 ERC and 40 NERC subjects.

Figure 1. Example of variations of the PIDS across TZ and PZ. A high PIDS value of 40% can be seen in the TZ region (a) and (b), while in PZ region the PIDS is 9.6%. Similarly, in another subject we can see a higher value of PIDS (33%) in PZ - (c) and (d), compared to TZ.

Figure 1. MRS prostate spectra from a volunteer in the frequency range 1.5 ppm to 5.5 ppm. a. Spectrum obtained with water signal suppression b. Spectrum obtained without water signal suppression but with water signal removal using a Löwner filter, c. Spectrum without water signal suppression but with water signal removal using a HLSVD filter.

Table 2. Mean values of the absolute metabolite concentrations of citrate and choline. The bottom row presents absolute metabolite tissue concentrations reported in reference 7. The choline values seem to be lower compared to the literature, but this is explained by the fact that the volunteers participated in the study were of a young age.

Wave propagation map, MRE (0-8 kPa) and MRE confidence map (0-8 kPa).

Left: SE-MRE 60 Hz. Right: GRE-MRE 60 Hz.

Figure 2. Distortion-free DIADEM-MRE magnitude image and SII-derived OSS map from the same healthy volunteer.

Figure 1: Standard EPI-MRE magnitude image, distortion-free DIADEM-MRE magnitude image, and T2-weighted anatomical image from a healthy volunteer

Figure 1 T2WI（A）image, DWI（B）image, fused APT and DWI (C, D) images of a BPH patient. T2WI（E）image, DWI（F）image, fused APT and DWI (G, H) images of a PCa patient. The placement of ROIs is as illustrated.

Table 2.Comparisonof APT and ADC values

Table 1. Scan parameters of APTw sequence in group A with the default parameters and group B with the optimal parameters

Figure2. a 75-year-old male with BPH of group B. T2WI(2a), DWI (2b), M0 (2c) (central zone SNR=71.22, peripheral zone SNR=22.12), APTw (2d) images are shown.

Figure 1 The middle slice for the whole prostate, apex and base of a randomly selected case segmented (peripheral zone (PZ) and transition zone (TZ)) by different approaches for scan 1 and 2.

Figure 3 The single score intra-class correlation coefficient (ICC) of the shape features extracted from the whole prostate gland (WP), peripheral zone (PZ) and transition zone (TZ) for the investigated methods. The methods connected with (-) were significantly different. In both manual and DL-based segmentations, Elongation, Flatness and Sphericity had a remarkably lower ICC than the other features in WP and TZ.

Figure 2. Spatial entropy maps for human and computer observer populations in 8 example patients. Spatial entropy maps were computed by first grouping the predictions from DL models trained with the same loss function. Four loss functions were investigated including crossentropy, DSC, Jaccard, and focal.

Figure 3. Autocontouring predictions (prostate: thick white line, OARs: blue lines) displayed with spatial entropy maps (transparent regions; yellow–30% of max entropy, orange-70% of max entropy, red-90% of max entropy) overlaid on post-implant prostate MRIs in a commercial treatment planning system. Arrows indicate regions of high spatial entropy clusters at organ junctions where physicians can review and potentially refine the automated predictions.

Figure 1. The schematic illustration of the meta-learning-based zonal segmentation network combines a 2D U-Net and an adversarial network for determining the shape prior.

Figure 2. Validation result of meta-learning based zonal segmentation model with the adversarial network, where (a) presents the T2w images, (b) shows the ground truth of zonal label, and (c) states the results of our approach. The CZ is masked in green, and PZ is in red.

The overall architecture of the proposed MC-DSCN network for prostate segmentation and classification. The coarse segmentation component (CSC) is constructed to generate coarse masks. The classification component extracts and fuses multi-parametric feature information and produce the lesion localization maps (CRM). The T2w images are concatenated with the CRM, then fed into the fine segmentation component (FSC) to generate the fine-mask.

(a-d)The comparisons for different segmentation network, including Unet, Res-Unet, coarse segmentation network(CSC), fine segmentation network (FSC). The loss is indicated in brackets, including dice loss and the fusion loss(dice loss, the online rank loss). Both CSC and FSC are based on a residual U-net with an attention block. (e-f) the comparison of PCa Classification ROC curve based on different inputs. (e) the inputs include only T2w, T2w with coarse-mask, only ADC, ADC with coarse-mask. (f) the inputs include ADC with coarse-mask, multi-parametric images with coarse-mask.

Figure 1. The proposed framework. Source domain labeled training data are employed first to train a network that generates pseudo labels for the target domain unlabeled training data. Then, the combined source domain labeled training data and target domain pseudo-labeled training data are utilized to achieve our proposed cross-domain cross-network optimization.

Figure 4. Example segmentation maps when transferring models from Domain 1 to Domain 2. From left to right, the three columns correspond to the transverse plane, the sagittal plane, and the coronal plane. From the head to bottom, the four rows refer to the ground-truth manual segmentation, the outputs of models trained directly on Domain 2 training set, the outputs of models trained on Domain 1 labeled training set, and the outputs of models trained on Domain 1 labeled training set and Domain 2 pseudo-labeled training set with the proposed cross-domain cross-network optimization method.

Figure 1: Segmentation examples of individual models: first column (A) shows the original image, second (B), third (C) and forth column (D) show masks of prostate, central and peripheral zones, respectively. Rows from top to bottom represents mag, c, phi maps, T2w, DWI, and ADC respectively.

Figure 2: Examples of segmentation results of the unified model: All segmented masks resulted from combining all mre maps as input, and the resulting masks were propagated to all other registered sequences. Columns A, B and C show masks of the prostate, CZ and PZ respectively. Top row is the mag images together with overlaid ground truth masks. The second row at the top to bottom show the predicted masks overlaid on mag, c, phi, T2w, DWI and ADC images respectively.

Figure 1 Phases of the machine learning challenge. 1) Data, Pre-processing and Feature extraction has been done. 2) Participant need to pre-process feature data with method of choice, including imputation of missing data, build a classifier with selected features, and validate the results with data from Site I. 3) External validation with unseen data from centers 1 and 2, with a limited number of evaluations. 4) The final test of the trained model and machine learning process with unseen data from centers 1 and 2, and completely unseen centers III and IV.

Figure 2 Data management of the machine learning (ML) challenge. Radiomic feature values and ground truth values are both given for Training/Validation of ML model (A). Only radiomic feature values are given for Leaderboard evaluations (B), where results are given five times to allow minor edits to the ML, and in final evaluation (D), where only one submission is done (C). Leaderboard and Test set (B-D) ground truth labels are hidden inside the Codalab server where performance evaluations take place by running python code through Docker package.

Fig. 1: Three cases of PI-QUAL score.

From left to right for each row: axial T2-weighted, diffusion (b-values and ADC map) and dynamic-contrast enhanced imaging.

PI-QUAL 2: only one sequence is of acceptable diagnostic quality (T2-WI). It is not possible to rule in and to rule out all significant lesions.

PI-QUAL 3: two sequences taken together are of acceptable diagnostic quality. It is possible to rule in but not to rule out all significant lesions.

PI-QUAL 4: at least two sequences are independently of diagnostic quality, and it is possible to rule in and rule out all significant lesions.

Table 1: The PI-QUAL score.

Figure 2. 3D deep residual neural network architecture for prostate cancer detection in this study. (a) R3D_PZ for prostate cancer detection in PZ. (b) R3D_TZ for prostate cancer detection in TZ. (c) Multi-task R3D (R3D_MT) for prostate cancer detection in both PZ and TZ. The input is a 32×32×11 spatial-b-value volume. Each “3D Conv” block consists of two 3D convolutional residual layers. The black arrows represent the shortcut connection.

Figure 3. Mean AUC, sensitivity, and specificity as a function of epoch number for PZ cancer detection (a, b, and c) and TZ cancer detection (d, e, and f) on testing data. The blue curve represents multi-task R3D learning, R3D_MT, the red curve represents R3D learning, R3D_PZ, in the first row or R3D_TZ in the second row.

Figure 1: ROC plot for a) LG vs. IG vs. HG and b) PI-RADS grade 4 vs. grade 5 classification using T2WI. Red curves stand for the performance of linear SVM, green curves for Gaussian SVM and blue curves for KNN classifier. ROC = Receiver-operating characteristics, CV = cross-validation, SVM = Support vector machine, KNN = K-nearest neighbour, DWI = Diffusion-weighted imaging, ADC = Apparent diffusion coefficient

Table 1: Features extracted from different texture models. FOS = First order statistics, GLCM = Gray level co-occurrence matrix, GLRLM = Gray level run length matrix, SFM = Statistical feature matrix, LTEM = Law's texture energy measures

The flowchart of the study. PCa: prostate cancer; BPH: benign prostatic hyperplasia.

Artificial intelligence architecture. a. Small Vnet (sVnet), based on V-Net with modifications, composed of one input block, four downsampling blocks (Down Block), four upsampling blocks (Up Block), four merge blocks (Merge Block) and one output block. The horizontal arrow denotes the transfer of residual information from the early stage to the later stage. b. Multi-small Vnet (msVnet), where two sVnets are cascaded.

Heatmaps of ICCs in (a) WP and (b) L datasets with selected good features (ICC ≥ 0.75) explain the patterns of repeatability in different prostate regions and lesions relative to pre-processing configurations and radiomic features.

The highest repeatability results on each prostate regions and lesions based on radiomic feature groups based on the and pre-processing configurations.

Figure 1. Representative T2 maps of ME-TSE T2 mapping with SENSE, C-SENSE and CS-AI. CS-AI significantly cleaned up the image noise.

Figure 3. Comparison of high-resolution T2 weighted image with CS-AI and proposed 0.7mm high-resolution ME-TSE T2 map with CS-AI.

Figure 1: Prostate cancer delineations of three radiologists (R1(red), R2(yellow) and R3(green)) on T2W and ADC. For smaller lesions (patient 1), all 3 readers had a good overlap. For larger (patient 2) and multi-focal lesions (patient 3), there was considerable variation in delineations that affect radiomics and robustness of classifiers.

Table 3: Individual and inter-reader performance assessment

Figure 3. ROC curves of the 4 radiomics models and TRUS biopsy in the training cohort (a) and test cohort (b). (Model 1 is based on manual segmentation of the prostate gland. Model 2 is based on manual segmentation of prostate cancer lesions. Model 3 is based on automatic segmentation of the prostate gland by the 3D prostate segmentation algorithm. Model 4 is based on thresholding segmentation of prostate cancer lesions by a fast-automatic thresholding algorithm. ROC: receiver operating characteristic; TRUS: transrectal ultrasound.)

Figure 5. Decision curve analysis comparing the net benefits of different radiomics models and TURS biopsy for the test cohort. (Model 1 is based on manual segmentation of the prostate gland. Model 2 is based on manual segmentation of prostate cancer lesions. Model 3 is based on automatic segmentation of the prostate gland by the 3D prostate segmentation algorithm. Model 4 is based on thresholding segmentation of prostate cancer lesions by a fast-automatic thresholding algorithm. TRUS: transrectal ultrasound.)

Figure 1. Flowchart of Gleason score classification. We classify the predefined lesion ROIs on the VERDICT maps using DenseNet and SE-ResNet. The network then gives the corresponding Gleason score to the lesion.

Figure 2. Results of five-point GS classification using two different networks: DenseNet and SE-ResNet on VERDICT. SE-ResNet with VERDICT generally yields better performance than DenseNet. VERDICT gives higher classification metrics compared to those reported in GS classification with bi-and multi-parametric MRI.

Fig 1: Information from a) histopathology, mpMRI images b) T2W, c) ADC, d) FA, and e) DCE) was combined in a logistic regression model to generate the f) cancer risk .aps in the TZ and the PZ. In this example, a combination in an ROI of hypointense T2W and ADC, hyperintense DCE resulted in a high-risk region in the cancer maps.

Fig. 2: Prostate cancer volume comparison of MRI cancer map vs histopathology. Cases with overestimated (underestimated) cancer volume are above (below) the solid blue one-to-one line. Dashed bounding lines were defined to indicate outliers (1cc ± 1.2·volume_pathology). 80% of cases were within the boundaries.

Figure 4. Coronal image results for subject 1 (top row) and subject 2 (bottom row). Soft-sense reconstructions are shown in a) and d). VN reconstructions are shown in b) and e) while the fully sampled, ground-truth images are shown in c) and f). The value indicated on the images is the calculated SSIM of the slice shown.

Figure 1. Data processing pipeline, and structure of the reconstruction network. First, a zero-filled reconstruction is generated from the under-sampled k space and two sets of coil sensitivity maps. Two sets of sensitivity estimates are required because the anatomy extends beyond the field of view, as described in Uecker et al.⁶ Then, this reconstruction, along with the raw k-space data and sensitivity maps are passed as the input to the reconstruction network. The reconstruction network has 10 stages (t =10), each stage applies data consistency and regularization.

Applying the pix2pix model to real measured 7T MRI data over different anatomies (left) and its corresponding output (right). Note that instead of showing the model input (8 single-channel images), the figure shows the sum of magnitude of these input images.

Graphical representation of data creation pipeline. Going from left to right, we start with a set of simulated B1+ and B1- fields, and a target prostate image. The B1+ and B1- fields are being registered to the shape of the prostate data whereafter the registered B1+ data is shimmed and scaled to mimic the signal of a T2w acquisition. Now combining the scaled B1+, B1- and prostate image results in the 8-channel input data for our model.

Fig.1 The overview of our framework.

Fig.2 A 61-year-old man with prostatic cancer with prostate specific antigen level of 50.88 ng/ml. Both of s-ADCb1000 and s-ADCb1500 had good performance in displaying the prostate, pelvic floor musclesm and pubic symphysis while s-ADCb50 produced blurry images of these structures. Regarding level of detail, s-ADCb1000 is more similar to z-ADC than s-ADCb1500 and maintains the details seen in z-ADC (red arrows).