Figure 2: Demonstrating the impact of DL processing. Four phases of a DCE scan are shown with arrows highlighting areas where DL delivers improved performance. Best viewed on a high brightness setting on the monitor

Figure 1: A) Offline trained spatial DL processing in which streaks are suppressed in the spatial dimension, B) Online trained temporal DL processing where a network is trained on the fly for suppressing streaks in the time domain, C) Flow chart of the overall workflow of the proposed method

Figure 3. The examples of extracted fetal pose of two different subjects in the dataset and the plots of the knee and elbow angles and the corresponding angular velocities.

Figure 4. Scatter plots of the three motion metrics vs. gestational age.

Figure 5: A representative case of the automatic segmentation

a: T2-weighted image

b: diffusion-weighted image (b=1000 s/mm²)

c: apparent diffusion coefficient map

d: A result of automatic segmentation of endometrial cancer overlaid on T2-weighted image

The tumor was well segmented despite the presence of hematometra (a:*) (Dice similarity coefficient=0.808).

Figure 3: The segmentation accuracy of our model for the test datasets with each MRI sequences as input data

Data are presented with mean±standard deviation

T2WI: T2-weighted image

DWI: diffusion-weighted image

ADC: apparent diffusion coefficient

Multi: T2WI, DWI, and ADC map

DSC: Dice similarity coefficient

PPV: positive predictive value

NPV: negative predictive value

The segmentation produced for a single representative slice from the validation dataset, showing the results from training without any augmentation (left) versus training with augmentation via rotation. Comparison shows both a reduction in false positive, or improperly segmented pixels, and an increase in true positive, or properly-segmented pixels.

The geometric transformations applied to the original dataset to produce augmented data. Each original scan was subject to each type of transform twice, with different parameters. Rotation was up to ±30° about the physical z-axis. Translation was up to ±30 pixels along the physical x- or y-axes. Reflection was across either the physical x- or y-axes. Rescaling brought the placenta up to 500x500 pixels, or down to 150x150 pixels from the original 256x256 scan. Operations were performed using Matlab⁶.

Figure 5. Accuracy of 3D liver localization and axial prescription, including the 6 edges: right (R), left (L), posterior (P), anterior (A), inferior (I), superior (S). Positive (negative) values, ie: purple (green) shading, indicate missing (excess) volume in the automated localization. Whole-liver coronal and sagittal prescription accuracy is similar to axial (not shown due to the additional considerations for coronal/sagittal S/I coverage). S/I coverage shows better accuracy than R/L coverage, possibly due to lower signal in the left lobe and complex anatomy.

Figure 3. Detection of the liver in each localizer orientation across several patients. a) In most cases, the liver volume was covered accurately by automated prescription. b) Missed coverage in the tip of the lateral segment of the left lobe due to insufficient axial localizer slices. The automated prescription aligned well with the manual annotation in patients with focal lesions (c), ascites (d), and/or cirrhosis (e). f) In patients with splenomegaly, the spleen abuts the liver and its signal level is close to that of liver, but the proposed method was able to select only the liver.

Figure 3: First two columns: eight repetitions of a DW liver slice (b = 800 s/mm²) with manually assigned labels. Last two columns: corresponding weight maps produced by the network after processing all patches using a sliding window. Locations with lower weight approximately coincide with image regions affected by signal-dropouts.

Figure 4: Qualitative analysis of uniform and CNN-based adaptive averaging, compared to the reference image obtained from averaging clean repetitions only. Using the repetitions and weight maps from Figure 3, the proposed method is able to recover signal which leads to higher agreement with the reference as confirmed by the difference maps (5x) as well as the normalized root-mean-squared error (NRMSE) and structural similarity (SSIM). Accordingly, elevated ADC values in the affected liver lobe (ROI, yellow box) are corrected with the proposed method.

Table 1. The results of the ROC analyses of the texture features extracted from QSM between cohorts of significant and non-significant hepatic fibrosis patients.

Only the results with statistically significant differences between the cohorts of significant and non-significant hepatic fibrosis are shown. AUC = area under receiver operating characteristic curve, SS = sensitivity, SP = specificity. DifVarnc = difference of variance, DifEntrp = difference of entropy, AngScMom = angular second moment, InvDfMom = inverse different moment.

Fig 1. The second-order texture analysis of susceptibility maps between cohorts of significant and non-significant hepatic fibrosis. The vertical axes are dimensionless quantities which are calculated from the gray level co-occurrence matrix. * indicates p < 0.05; ** indicates p < 0.01. DifVarnc = difference of variance, DifEntrp = difference of entropy, AngScMom = angular second moment, InvDfMom = inverse different moment.

Figure 2. The framework of the proposed multimodal fusion method.

Figure 1. Contrast-enhanced MR images of a 59-year-old man with pathologically confirmed low-grade HCCs (grade II, white arrow) (a) the pre-contrast phase (b) the arterial phase (c) the portal vein phase (d) the delayed phase.

Figure 1. Overall framework of the proposed PM-DL model, including a) image registration and liver segmentation, b) screening of suspicious lesions in DWI images, and c) identification and segmentation of true lesions in DCE images.

Figure 4. Two representative sHCCs correctly detected by the PM-DL model in the external test cohort, a) LR-4 according to LI-RADS v2018, which can be regarded as HCCs, Edmondson–Steiner grade III; b) LR-M according to LI-RADS v2018, which cannot be regarded as HCCs, Edmondson–Steiner grade II. (Yellow box: extracted image patch, green contour: manually labeled mask, yellow contour: model predicted mask).

Flowchart of study population.

Table 1. The discrimination of single sequences in predicting 2-year recurrence with logistic regression.

Figure 1. 71-year old patient with intraductal papillary mucinous neoplasm. (First line from L to R: MRCP_{FAST 3D} with AiCE, MRCP_{Compressed SPEEDER} with AiCE and MRCP_{conventional SPEEDER} with AiCE; second line from L to R: MRCP_{FAST 3D} without AiCE, MRCP_{Compressed SPEEDER} without AiCE and MRCP_{conventional SPEEDER} without AiCE).

MRCP_{FAST 3D} and MRCP_{conventional SPEEDER} more clearly depict intrahepatic bile duct as well as main pancreatic duct than MRCP_{Compressed SPEEDE}R, when applied AiCE or not. In addition, AiCE was able to improve image quality of MRCP obtained by each technique.

Figure 4. Compared results of interobserver agreement and qualitative image qualities among all methods.

Interobserver agreement on each method was ranged between 0.44 and 0.77. When applied AiCE, each index on MRCP obtained by each technique and reconstructed with AiCE were significantly higher than those without AiCE (p<0.05). On each index, MRCP_{Compressed SPEEDER} with and without AiCE were significantly lower than MRCP_{FAST 3D} or MRCP_{conventional SPEEDER} with and without AiCE (p<0.05).

Figure1. Receiver operating characteristic curves (ROC) of the training (A) and validation (B) cohort. AUC, area under the receiver operating characteristic curve.

Table 1. Radiomics features extracted based on enhanced MRI

Figure 1: GradCAM visualisations for two false positive classifications

Figure 2: GradCAM visualisations for two true positive classifications

Figure 3: Visualization of segmentation results

Figure 1: Overall framework

Table 1. The architecture of ResNet-50

Figure 2. The feature extraction process of ResNet-50.

Figure 1. Network structure of the deep learning-based fast MR reconstruction framework used in this study.

Figure 4. Example of lesion analysis.

Flow diagram of network training with folded image training strategy (FITS) and image reconstruction. In training with FITS, images for training were folded by factor of 2 to reduce memory consumption.

A representative case. (CS) total variation regularized compressed sensing showed some aliasing on liver parenchyma. A few aliasing is remained in the conventional model-based deep learning reconstruction (conv-MoDL), while it is removed in the model-based deep learning reconstruction using FITS (FITS-MoDL).

Figure 2. a. ROC curve analysis of radiomics models in the testing cohort (AUC=0.85). b. ROC curve analysis of Fibroscan(AUC=0.83), APRI(AUC=0.68), FIB-4(AUC=0.69).

Table 1.The eighteen valuable features.

Figure 2. Comparison between two results (Seg1 vs. Seg2) of Dice coefficients using the different trained models. A.) Dice coefficient plots from a model trained using one subject (1Subj); B.) using three subjects (3Subj); C.) using the six subjects (6Subj). D.) the bar plots for the above three cases. Seg1 indicates the results predicted using the first network alone; Seg2 indicates the results by using the two networks in Fig. 1. Gray areas in A-C indicate the training data sets.

Figure 1. Diagram of two neural networks including UResNet1 and UResnet2. The inputs for UResNet1 are three consecutive slices; the inputs for UResNet2 are three consecutive slices and one predicted mask for the third slice. Seg1 indicates the output of the first network; Seg2 indicates the output of the whole networks. The dotted and dashed lines indicate the relationship in the prediction stage.

Figure 1. Flow chart of experiment design

Figure 4. Comparison between renal layer’s T2* and measured renal size changes.

Fig. 2 Box-plots of IVIM-DWI Parameters.

Panels A-D represent ADC, D, D* and f values for G-1 (cervical squamous cell carcinoma), G-2 (cervical adenocarcinoma) and G-3 (cervical small cell carcinoma), respectively (p ＜ 0.05).

Fig. 3 ROC Curves of Regression Model for G-1 Compared to G-2, and Individual Independent factors (D, D* and Gray Level Variance).

Figure 3. Examples of the comparison between CNN-predicted and manual segmentation. (a) Section examples of eight bones on DWI image; (b) The corresponding overlapping images between manual segmentation (white background) and CNN-predicted segmentation; (c) Section examples of eight bones on ADC image; (d) The corresponding overlapping images between manual segmentation (white background) and CNN-predicted segmentation .CNN: Convolution neural network.

Figure 2. The SCORE system and evaluation criteria on DWI and ADC images. Condition A refers to that the location of the predicted CD is consistent with manual CD, and the range of the predicted CD is larger than (A1) or less than (A2) the manual CD, or partially overlaps with the manual CD (A3). CD: connected domain, which is defined as the label with a continuous structure in 3D space. DSC: Dice similarity coefficient.

Figure 4: FSE and PROPELLER images reconstructed using deep-learning based reconstruction methods. Deep-learning based reconstruction method improved the SNR and in-plane resolution of FSE images, but the motion artifacts were not removed. Due to the robustness to motion and deep-learning based reconstruction method, the PROPELLER images showed better SNR, in-plane resolution and less motion artifacts.

Figure 5: High resolution rectal FSE and PROPELLER images of a patient. Deep learning reconstruction method improved the SNR and in-plane resolution. Due to the robustness to motion, PROPELLER further improved the image sharpness.

Figure 1. 51-year old patient with prostatic cancer in the left transitional zone

b: DLR improves image quality of DWIs for each b value. Each DWI shows the prostatic cancer as high signal intensity in the left transitional zone (arrow).

Figure 1. 51-year old patient with prostatic cancer in the left transitional zone

c: All ADC maps show the prostatic cancer as a hypointense signal in the left transitional zone (arrow).

Figure 1. The conventional T2-weighted FSE images and PROPELLER images without and with DL are shown. Because the phase encoding direction of AP was often chosen to avoid aliasing and long scan time, conventional T2-weighted FSE is sensitive to motion artifacts in sagittal imaging (A). PROPELLER is robust to motion and minimized the motion artifacts (B), however, the SNR and in-plane resolution were limited by the clinical scan time. DL reconstruction improved the SNR and in-plane resolution without increasing scan time (C).

Figure 3. This graph shows the scoring of the best sequence on the 4 images by three radiologists. DL 50 and DL 75 were considered to be better than Non-DL and DL25.

Table 1: MR enterography protocol before and after intervention

Table 2: The intervention led to an overall decrease in exam length and standard deviation, with a resulting increase in efficiency.

Figure 1:A 56 year old patient with cervical mucinous adenocarcinoma.1A shows T2WI image and the red arrow points to cervical cancer;1B shows the MK pseudo color map of DKI sequence of corresponding layer; 1C shows the ROI outline of corresponding layer, and the red area shows the coverage area of tumor parenchyma.

The ROC curve for LR classifier

Fig. 4 AUCs for predicting pCR using logistic regression model at T0 (top) and T1 (bottom). The number of patients and pCR rate included in analysis at T0 were: 941 (35%), 368 (18%), 149 (39%), 81 (68%), 343 (43%) for full, HR+/HER2-, HR+/HER2+, HR-/HER2+, HR-/HER2- respectively. The numbers at T1 were: 904 (34%), 357 (18%), 142 (39%), 74 (66%), 331 (42%). FTV – functional tumor volume, HR – hormone receptor, HER2 – human epidermal growth factor receptor 2.

Fig. 3 Correlation map among FTV and shape features at T0.

(A)ROC curves for nomogram, radiomic signature, age and meninges predicting OS in validation set. (B) The stratification performance of the nomogram in validation set.

(A) and (B) show the ROC curves of the nomograms in the training and test sets; (C) and (D) show the calibration curves of the nomograms in the training and test sets; and (E) and (F) show the DCA curve of the nomograms in the training and test sets.