Representative image registration results of our proposed method. Our proposed methods are robust to the inconsistent contrast between the moving image and the reference image, whereas such inconsistency tends to cause distortions in ANTs results.

(a) Proposed framework for motion correction in cardiac MRI. The image registration neural network inputs are segmentation network outputs of the reference image and the moving image. The output of the registration network is the deformation field that is applied to the moving image to yield the registered image. (b) Registration neural network architecture.

Figure 2. Overview of the proposed self-supervised denoising framework.

Figure 1. Visualization of data model for cardiac MRI denoising.

Fig 2. Illustration of the proposed method, including the multi-task network aiming joint regression and classification.

Fig 4. A comparison of 3D activation map reconstructed by (a) manual label; (b) active contour; (c) regression network; and (d) our method.

Figure 4. Results along different directions. Representative axial (column i), coronal (column iii), and sagittal (column v) image slices and enlarged regions (columns ii, iv and vi) from different methods including standard (row a), Wave-CAIPI (row b), HDnGAN (λ=0) (row c), HDnGAN (λ=10^-3) (row d) of 3 evaluation subjects.

Figure 2. Effects of the adversarial loss on image quality. Representative axial image slices (rows a and c) and enlarged regions (rows b and d) from different methods and weights (λ) of the adversarial loss of a multiple sclerosis patient. Metrics including the mean absolute error (MAE), peak signal-to-noise ratio (PSNR), structural similarity index (SSIM), and VGG perceptual loss are listed to quantify the similarity between images from different methods and the standard FLAIR image.

Figure 4. Segmentations from different models. For each model, the 4 best (left) and worst (right) segmentations are shown with Dice score/SNR (these were calculated from the whole 3D volume). The models’ segmentations are shown in red, those from the expert readers in blue, and their overlaps in purple. White arrows indicate areas where models deviated from ground truth. From this comparison, the model trained with registered data performs better than unprocessed data but struggles with low-SNR and poor ventilation images. Our proposed method (model 3 & 4) helps solve this problem.

Figure 5. Box plot and table of Dice score for each model tested with the “pristine” test dataset. The orange lines indicate the median of the score while the dash green lines indicate the mean. By comparing model 1 and 2, the plot shows the improvement in using the registered dataset for training 3D-CNN but with a drawback shown by low-score outliers due to poor ventilation and low-SNR images. Our proposed method (model 3 & 4) solves the problem and eliminates these outliers.

Figure 4. Robustness to moving-image MRI contrast across 10 realistic A spoiled gradient-echo and B MPRAGE image pairs. In each cross-subject registration, the fixed image has consistent T1w contrast. Towards the right, the T1-contrast weighting of the moving image increases. Our method (sm) remains robust while most others break down, including VoxelMorph (vm). As errors are comparable across methods, we show error bars for ANTs only and indicate the standard error of the mean over subjects.

Figure 3. Image synthesis from random shapes (top) or, if available, from brain segmentations (bottom). We generate a pair of label maps {s_m,s_f} and from them images {m,f} with arbitrary contrast. The registration network then predicts the displacement u. In practice, we generate {s_m,s_f} from separate subjects if anatomical labels maps are used. We emphasize that no acquired images are involved.

Proposed I-DSLR-SEG-E2E network architecture. A K-iteration I-DSLR network is cascaded with a CNN for segmentation. It is trained end-to-end.

Comparison of reconstruction and segmentation quality of various methods on 6-fold undersampled k-space measurements. Reconstruction SNR in dB along with dice coefficients for CSF, GM and WM are reported for the particular slice. The methods in red box typically cascade separately trained tasks and the blue one is the proposed end-to-end training approach.

Figure 3: Pipeline of our deep learning framework. Preprocessing of the data consists of organ segmentation, followed by data normalization and creation of TFRecords for efficient data processing, which are then split into training and test data and resized such that all images have the same size. The modular design allows to choose arbitrary deep learning models to investigate different network architectures. Dashed lines mark optional modules (data augmentation, transfer learning). As data augmentation, we implemented random rotation, random shift and elastic deformation⁹.

Figure 1: Exemplary abdominal slices of one examination: images with fat (a) and water weighted (b) Dixon sequences, ADC map (c) and PET image (d) were acquired. For each patient, examinations have been conducted prior to, two weeks and two months after starting immunotherapy.

Figure 1: Training procedure of the GAN. The generator consists of a U-Net architecture of residual blocks with adaptive instance normalization layers that inject the input acquisition parameters (y_g) in the encoder part of the generator and the output target labels (y_t) in the decoder part of the generator.

Figure 2: Example pair of ground truth input image g with its labels y_g and the corresponding fat-saturated target image t with its labels y_t. The generator is trained to predict the target contrast from the input image and the corresponding input and target acquisition parameters: G(g, y_g, y_t). The last image shows the absolute error map of target and prediction. The images are annotated with the real acquisition parameters TR and TE and the acquisition parameters as predicted by the AC for G(g, y_g, y_t) .

Figure 1. The ARN (CNN + FCN) architecture for image registration. The input layer took the moving image (128x128x40 matrix) as an input, 20 channels were used for the 5 CNN layers. Six parameters that contain x, y, z shifts, and x, y, z rotations were generated after the 2 FCN layers. Bilinear interpolation was used to obtain the moved image by applying 6 parameters to the moving image. The ARN was trained to minimize the loss function, which included the MSE loss, pixel-wise L1 loss, and SSIM loss between the fixed and moved images.

Figure 2. Comparison of loss between ARN and SPM registration. (a) MSE loss, (b) Pixel-wise L1 loss, (c) SSIM loss using simulated perfusion data, and (d)MSE loss, (e) Pixel-wise L1 loss, (f) SSIM loss using real perfusion data were compared between ARN and SPM registration.

Figure 2: SGL accurately predicts ALS. Left: classification probabilities for each subject’s ALS diagnosis. Controls are on the left, patients are on the right, predicted controls are in blue, and predicted patients are in orange. The SGL algorithm achieves 86% accuracy. Right: SGL coefficients are presented on a skeleton of the major white matter fiber tracts. The brain is oriented with the right hemisphere to our left and anterior out of the page. As expected large coefficients are in the fractional anisotropy of the corticospinal tract.

Figure 4: Predicting age in the Healthy Brain Network, a large pediatric neuroimaging study. Left: The predicted age of each individual (on the abscissa) and true age (on the ordinate), from the test splits (i.e., when each subject’s data was held out in fitting the model); an accurate prediction falls close to the $$$y = x$$$ line (dashed). The mean absolute error in this case is 1.5 years and R² = 0.56. Right: Standardized residuals (on the abscissa) as a function of the true age (on the ordinate). Predictions are generally more accurate for younger individuals.

Figure 1: Block diagram of the proposed method: A Deep k-means Based Tissue Extraction from the reconstructed Human Brain MR Image. (A) shows the deep learning approach for image reconstruction and (B) shows the k-means clustering algorithm used for the brain tissue extraction.

Figure 3: Results obtained from the proposed Deep k-means and CG-SENSE k-means: (A) shows the reconstruction results obtained from U-Net and CG-SENSE, (B) shows the segmentation results obtained from k-means clustering algorithm. (C-E) show the extracted white matter, gray matter and CSF from the proposed method and CG-SENSE k-means.

Fig 3. (a) Anomalous MOOD validation data (b) Model does not reconstruct the anomaly in the output (c) Mask representing the difference between the input image and the reconstructed image, indicating the presence of an anomaly or out of distribution sample in the input image. By morphological closing along with binary thresholding, the edges of the anomaly are detected and hence localized

Fig 2. (a) Non-anomalous MOOD data which is the input to our model (b) Reconstructed image obtained as the output of our model (c) Mask representing the difference between the input image and the reconstructed image, with 0 value indicating no anomaly or out of distribution sample in the input image.

Fig.4: Maximum intensity projection and a montage of slices for the Interpretability result using Integrated Gradients, overlaid on the input volume. White pixels were converted to red for overlaying

Fig.5: Maximum intensity projection of layer-wise activations, generated using Guided Backpropogation. From left to right, top to bottom: initial layer of the network to the final layer, and finally the output from the model

Fig.1: The image shows a representative examples of D, f and D* maps for each SNRs, in simulations, generated with our neural network model and their respective ground-truth images.

Fig.4: The image shows in-vivo IVIM maps examples generated using: our neural network a), Bayesian method b), Barbieri's neural network c), and Bertleff's neural network d).

Figure 4. Sample 2AFC trial for network with 64 channels and 0.3 dropout including artifacts which do not affect the detection task. The artifacts may be hyper-enhanced features of the true image.

Table 1. Results for the combinations of initial number of channels and amount of dropout. The choice that did consistently well across all metrics we considered was 64 channels with 0.1 dropout but the network with 64 channels and 0.3 dropout has very similar human observer performance. All networks perform similarly for human detection except for the network with 64 channels and 0 dropout. We also see that the approximation to the ideal observer (LG-CHO) performs similarly except for the networks with 32 channels and 0.3 dropout and 64 channels and 0 dropout.

Figure 1. The architecture of deep supervision U-net++. U-net++ consists of an encoder and decoder path that are connected with nested and dense skip connections.

Figure 2. The distribution of the Dice coefficient (DSC) values for the automatic segmentation results of CN, GP, PUT, RN, and SN regions.

Figure 3. Sample 2AFC trial where a subject chooses which of the two images contains the signal in the middle. Each of the 4 observers conducted 200 trials for each amount of undersampling.

Figure 2. U-Net Diagram. For this study x = 64 channels and a 0.1 dropout was used.

Fig. 4 Denoising low field T₁-weighted images: Training (a-c) - a) a representative image used as the gold standard for training; b) noise added image used for the training; c) the output of the native noise denoising network (NNDnet). The corresponding magnified images are shown on the left for the red square shown in a). Testing (d-g) - d) a representative 0.36T noisy image e) gradient anisotropy diffusion denoised (AD) result that is blurry; f) NNDnet denoised image; g) NNDnet + AD denoised image. The right column contains corresponding magnified images

Fig. 2 Denoising tailored MR Fingerprinting (TMRF): Training (a-c) - a) a T₁-weighted image from the human connectome database; b) extracted noise from the TMRF data added for training; c) corresponding native noise denoising network (NNDnet) result. The left column shows the corresponding magnified images for the red square shown in a). Testing (d-g) - d) a test TMRF T₁ image that suffers from noise e) corresponding gradient anisotropy diffusion denoised (GADD) result; f) NNDnet denoised image; g) NNDnet + GADD denoised image; corresponding magnified images on the right.

Figure 2 Three examples of the 3D lumen segmentation results with the manual labels as reference. An automatic skeletonization algorithm was used on the predicted binary segmentation to extract the centerline (Red lines) of the intracranial arteries.

Illustration of the proposed VoxelHop framework. Deformation fields within the whole brain and tongue derived from registration between all subjects and the atlas are input into our framework.

Comparison of the receiver operating characteristic curve between our SSL-based VoxelHop framework and the comparison methods including 3D VGG and 3D ResNet.

Fig. 5. Comparison of the four methods using the phase factor (phasor) images (First column) of (A) brain & neck (coronal); (B) breast (transverse); (C) hand (coronal); (D) upper abdomen (transverse); (E) lower abdomen and pelvis (coronal); (F) thighs (transverse); (G) calves (transverse); (H) ankle (sagittal). Red arrows indicate the unsolved regions. Color bars of the unwrapped phase images are shown on the right side.

Figure 1: the unsupervised network architecture of the proposed method

Figure 1. uU-Net architecture. Conv: convolution. IN: instance normalization. ReLU: rectified Linear Unit.

Figure 3. Representative pelvic MR images of three patients with generated overlays.

Figure 3. ANN prediction results from a single-shot spiral trajectory with 251 points for (A) tumor detection, (B) tumor size classification, and (C) tumor location classification.

Figure 2. The proposed ANN model structure for tumor detection, and size/location classification. The model input is a vector concatenated by the real and imaginary part of the complex k-space data, outputting three categories of probabilities. Three tasks share the first two fully connected layers (FC), and each FC is followed by a batch normalization layer, a rectified linear unit (ReLU) layer and a dropout layer (50%). Each task has its own fully connected layer and softmax layer for classification. Noted that tumor size/location is only classified in presence of tumors.

Training strategy. An axial slice of a representative image is shown unmasked (b.1) and masked (b.2). In all cases the label (or ground truth image) was the Laplacian PCG unwrapped phase, and the input was the rewrapped phase of the label such that the label was the exact unwrapping solution of the label. Differences can be seen between the raw and the rewrapped phase images.

The CNN unwrapping performance on raw phase images. A coronal, axial and sagittal slice of the raw phase (a) and phase unwrapping solutions using the masked CNN (b), Iterative Laplacian PCG (c) and SEGUE (d) for a representative healthy volunteer in vivo. The computation times are shown below. Red arrows indicate where the Laplacian or SEGUE solutions appear more accurate than the CNN. The yellow arrows indicate where the CNN was more accurate.

Figure 4: The resulting T map of two-sample T-test (AD vs NC). The top row shows the results obtained by DLASL. The bottom row shows the results obtained by the non-DL-based conventional processing method[5]. From left to right: slices 95, 100, 105, 110, 115, 120 and 125. Display window: 4-6. P-value threshold is 0.001

Figure 3: The box plot of the CNR (top row) and SNR (bottom row) from 21 AD subjects' CBF maps and 24 NC subjects' CBF maps with different processing methods.

FIGURE 4: Visual example of background phase-error correction. Coronal view of the aorta and common iliac arteries during peak systole, with flow velocity represented by a colormap ranging from red (80 cm/s) to blue (0 cm/s). For assessment of flow continuity, measurements (in L/min) were taken at multiple locations. Corrected velocity measurements demonstrated improved consistency along the length of the infrarenal aorta as well as conservation of mass post-bifurcation.

FIGURE 5: Bland-Altman analysis. Consistency of flow measurements are shown for (a) uncorrected velocity data, (b) velocity data corrected manually, and (c) velocity data corrected automatically. Light blue points represent comparisons of arterial and venous flow, while dark blue points represent comparisons of flow before and after vessel bifurcation. Corrected measurements in (b) and (c) demonstrate greater consistency with narrower limits of agreement than seen in (a).

Figure 2. Validation results. Field map augmentation modified the frequency range of each slice depending on the parameters 𝛼 and β. a) shows the different combinations for an example slice and the achieved frequency ranges. b) Image panel displaying the blurred image, gold standard, U-net correction with field map, U-net correction without field map, and MFI correction images.

Figure 4. Filter visualization. Visualization of representative filters from the 2nd, 5th, 22nd, and 25th convolutional layers for four different brain slices from the testing dataset for model explainability.

Figure 1. The overall workflow of the proposed task-driven MR imaging method, which consists of two key components (teacher forcing and re-weighted loss training schemes) and two modules (reconstruction and segmentation modules).

Figure 3. Visualization of the segmentation results of the proposed method compared with existing methods.

Architecture of Autoencoder with Feature Fusion. The encoder and decoder each have eight convolutional (conv) layers with pooling and batch normalization (BN). Each conv layer had a kernel size of 9 and 16 filters and one fully connected layer with 1000 hidden units. The feature maps of all the layers in the encoder are concatenated using maximum pooling and passed to the latent vector (Z). Z is a fully connected layer with 128 hidden units with a linear activation function. All the layers in the model have a ReLU activation function, except the last layer has a tanh activation function.

Correlation of metabolite concentration obtained using LCModel, for AE predicted and water suppressed spectra for the in-vivo test dataset. The x-axis shows the absolute concentration values estimated from water suppressed spectra and y-axis shows the corresponding AE predicted spectra. Overall, there is a strong correlation between metabolite for tNAA, tCho, tCr and Glx, Lac, GSH, mI and 2HG has a decent correlation. The clustering pattern seen in other metabolites can be caused by reconstruction error and change in the degree of freedom of LCModel fit between the spectra.

Figure 5: Force-ranking summary. DLR’s average force-ranking score is consistently higher than the other methods in all pairwise comparisons and in all anatomy groups. DLR was rated statistically higher than the 3 counterparts in 15/18 pairwise comparisons (p < 0.012) except the three instances annotated by as NS (non-significance).

Figure 4: Average readers’ scores for 6 anatomy groups. DLR’s average scores are consistently higher than those of the 3 other methods in 143/144 pairwise comparisons (6 anatomy group x 8 criteria x 3 pairwise comparisons) except 1 instance, where DLR’s average score is smaller than that of GA53 by a margin less than 1% (i.e. 4.63 for DLR vs. 4.67 for GA53). DLR is rated statistically higher all other three methods (p < 0.017) in 134/144 pairwise comparisons.

Figure 1. Federate Learning Workflow

Figure 3. ACC Comparison of Federated Trained Models with Baseline Models

Figure 1: Correlation between the inter-rater Dice score and uncertainty measures (pooled data from the validation and test datasets). The marked test cases are shown in Figure 2.

Figure 2: Selected axial slices from three cases (marked A, B, C in Figure 1). For each of the three cases, the left panel shows an axial slice of the T2W-FLAIR image with the segmentation labels. The right panel shows the corresponding uncertainty maps (brighter areas correspond to higher uncertainty) illustrating areas of high and low uncertainty of the segmentation model. Segmentation labels: turquoise: overlap between the labels of rater 1 and 2; magenta: rater 1 only; orange: rater 2 only.

Figure 1: A) Sagittal cube sequences evaluated for meniscal lesions (arrow pointing to post. horn tear in medial meniscus). B) Swarm platform interface used to derive consensus grades for location of lesion. C) Visualization of the trajectory of decision made by the swarm. While there were individually divergent opinions, the eventual consensus of the group in this example was for posterior horn of the medial meniscus.

Figure 2: Resident versus Ground truth (GT). A) Confusion matrix (CM) for 3 resident majority vote vs GT (kappa: 0.01) B) CM for 3 resident swarm vs GT. Accuracy improves compared to majority vote (kappa: 0.24) C) CM for 5 resident majority vote vs GT (kappa: 0.05) D) CM for 5 resident swarm vs GT. Accuracy improves compared to majority vote (kappa: 0.37). Note: 5 resident swarm was unable to obtain a consensus on 1 exam, which was excluded during CM tabulation.

Figure 1. T1 data of 25 subjects were scanned using two scanners. We refer to the T1 data from scanner 1 as dataset 1 and from scanner 2 as dataset 2. The data preprocessing steps included: 1) N4 bias field correction and skull-stripping using our in-house software (https://github.com/pnlbwh/luigi-pnlpipe); 2) Creating label maps using FreeSurfer v. 7.1.0; 3) Affine registration using antsRegistration.

Figure 2. The summary of our multi-site T1 data harmonization network: CycleGAN with segmentation loss

Figure 1. Sample ultra-low-dose protocol. The participants were scanned in two sessions and the two sets of MR images obtained from the sessions were used in separate neural network trainings to test the effect of simultaneous (S) vs. non-simultaneous (NS) inputs.

Figure 3. Representative amyloid PET images (top: amyloid negative, bottom: amyloid positive) with the corresponding T1 MR images. Both sets of CNN-enhanced ultra-low-dose PET images show greatly reduced noise compared to the ultra-low-dose PET image and resemble the standard-dose PET image. NS: non-simultaneous, S: simultaneous

Figure 2. Transverse (top), coronal (middle) and sagittal (bottom) views of T1-weighted MR Image, overlaid with ROIs by manual segmentation, MTL-generated segmentation, and CIC-generated segmentation. Transverse slices of manual segmentation show postPU to have slightly irregular shape compared to a smoother postPU in MTL-generated segmentation. Coronal and sagittal slices of the MRI and manual segmentation also show VST to have irregular shape compared to a smoothed and more circular VST in MTL-generated.

Figure 3. Scatter plots between BPND of the 5 subregions calculated using manual ROIs and both MTL- (left) and CIC-generated (right) ROIs across 19 independent test subjects. Regression line for all subregions is overlaid scatter plots, with regression coefficients found in Table 2.

Figure 2. Qualitative Results of two representative slices. (a) Input MR images, ground truth PET images, and the synthesized PET images by three methods. (b) Spatial attention map from Attention U-Net and the proposed method.

Figure 1. Overview of the proposed method.

Figure 4 Top row: PET image reconstructed using the 4-class umap, 4-class umap with bone information from CT and 4-class umap and corrected using the generated bias image from left to right. All images are scaled the same. Below are the relative error of each PET image compared to the PET image reconstructed using the CT enhanced umap.

Figure 3: Real bias images and bias images generated by the pix2pix network. The bias image values correspond to the real and generated relative error between the PET image reconstructed using the 4-class attenuation correction map and the PET image reconstructed using the 4-class attenuation correction map with added bone information from a co-registered CT image. The images are from the hold-out test set.

Figure 1: Illustration of the proposed multimodal fusion framework. The middle panel shows the tensor model-based joint spatial-modality-intensity distribution and spatial prior representations. The bottom panel shows the structure of the fusion network integrating the outputs from different classifiers and spatial atlases.

Figure 2: The performance comparison by adding in each component in the proposed fusion framework. The segmentation accuracy was improved with more features being added, which confirms the advantage of capturing intensity-modality-spatial dependence information

Fig. 2 Examples of SCCS and MCCS results for AF = 2. (a) Unet-DC. (b) MoDL. The number of epochs in the training were 1 for SCCS and 5 for MCCS.

Fig. 1 Experimental condition. (a) Acquisition parameters. (b, c) Input/output images for (b) SCCS and (c) MCCS. US-FLAIR: undersampled FLAIR, US-T1WI: undersampled T1-weighted image.

Fig.1. Overview of the method. Two images, including the moving FLAIR image(M) and the fixed DWI image(F), are input into D_θ to generate the transformation filed(φ). D_θ represent the registration network.

Fig.2.Qualitative registration results of different methods. The first two columns present the moving FLAIR images and the fixed DWI images with the annotated stroke lesions. The third columns show the registration results of our proposed method, and the rest columns are the results generated by the comparison algorithms.

Figure 1. Preprocessing of spatial factors and the proposed network architecture

Figure 2. An example case of synthetic images using different methods. (a) T1 MPRAGE, (b) T1 GRE, (c) T2 FLAIR.

Figure 2: Illustrations of background segmentation results for the development database.

The columns display segmented slices per modality and anatomical region, whereas the rows correspond to the 3 different visual quality ratings (Perfect, Acceptable, Not acceptable). Dice, Jaccard and AMI values for the 3D volume are reported under each slice. Only two DSC brain volumes segmentations were found not acceptable.Their segmentations are displayed with transparency to show the underlying tissue.

Figure 1: Background segmentation performances in the development database.

Top: Quantitative results with values of Dice coefficient (Dice), Jaccard similarity index (Jaccard), and adjusted mutual information (AMI). Dice and Jaccard values range between [0-1], 1 means the two segmentations are identical. AMI values are <=1, 1 means the two clusters are identical.

Bottom: Qualitative (visual) evaluation of obtained segmentations with 3-stage rating: perfect, acceptable (i.e. without tissue voxels removed) and not acceptable (i.e. with tissue voxels removed) results.

Figure 2 Multi-contrast reconstruction for multi-contrast MR data at R=3, clearly demonstrating the proposed model could reconstruct high-fidelity images without obvious artifacts.

Figure 1 The proposed method uses a Res-UNet architecture, which consists of 4 pooling layers. Real and imaginary parts of complex T1- and T2-weighted images are treated as separate channels for the model inputs.

Figure 2: WMH volume evolution for each subject according to age. Each subject was attributed a Fazekas score at baseline. This demonstrates the variable growth of the WMH for each subject even for those with an older age.

Figure 1: WMH volume per Fazekas score; both evaluated at baseline. WMH volume is highly correlated to the Fazekas score as shown by the clear separation of the different classes in the boxplot and the Spearman correlation of 0.921 (p-value < 0.001).

Fig. 1: A robust, automatic pipeline to quantify the stained area fraction (SAF) from histology slides as highlighted in the 4 steps. Input RGB slides are processed to produce SAF maps at variable resolution. We aim to correlate SAF maps at MRI-scale resolution (512x512 µm²/pixel) to MRI measures.

Fig. 3: An example SAF map (top row) and the absolute percent difference maps for all within-subject pairwise comparisons between adjacent slides with local thresholding (i.e. proposed pipeline). Each column shows a different subject. 3 adjacent slides produce 3 pairwise comparisons (subjects 6,7), while 4 adjacent slides produce 6 pairwise comparisons (other subjects). For almost all subjects, the highest percentage difference is found on the edges of the tissue, implying possible misalignment after co-registration or reduced robustness to tissue edge artefacts.

Figure 1: Schematic diagram of deep learning model and radiomics model applied.

Figure 2: Feature heatmaps of a representative patient generated from the ResNet50. (A) T2WI image (left) and the corresponding feature heatmap (right). (B) DWI image (left) and the corresponding feature heatmap (right).

Figure 1. Grad-CAM visualizations on tumor detection for different training networks. The top row depicts the original MR image examples from four subjects. The magenta counters indicate the tumor lesion boundaries. The bottom rows show the Grad-CAM visualizations for three different training algorithm on the selected axial slices.

Table 1. Mean classification and localization error (%) on the testing database for DenseNet, GoogleNet, and MobileNet.

Fig.3: Result from experiments with ICNet.

(a) Intramodal deformable registration of affinely registered pre-processed images, after 100 epochs.

(b) Intermodal deformable registration of affinely registered pre-processed images, after 200 epochs.

Fig.1: Result from experiments with VoxelMorph, after 1000 epochs.

Rows (top to bottom): fixed image, moving image, and warped image

Columns (left to right): Intermodal deformable registration of affinely registered pre-processed images, Intramodal deformable registration of affinely registered pre-processed images, and Intramodal registration performed directly on raw images without using any pre-processing.