Search

Perovskite‐oxide‐based ferroelectric tunnel junctions (FTJs) hold great potential for applications in non‐volatile memory and neuromorphic computing due to their unique properties. However, the challenges in synthesizing high crystalline quality perovskite oxides directly on silicon wafer limit the applications of these FTJs in conventional Si‐based integrated circuits, let alone the neural networks. Herein, perovskite oxide FTJs with an ON/OFF ratio up to 1.2×106, writing/erasing speed down to 1 nanosecond, and cycling endurance (>106) are achieved by integrating ultrathin freestanding ferroelectric perovskite oxide membranes directly on silicon wafers using a wet‐transfer method. Moreover, synapses based on these FTJs exhibit long‐term plasticity. For handwritten digits recognition task, the convolutional neural network (CNN) simulation is implemented achieving a recognition accuracy up to 98.9% based on the experimental performance, close to the result of 99.2% by software‐floating‐point‐based CNN. This work sheds light on the integration of ferroelectric perovskite oxides directly on silicon for high‐performance FTJ‐based non‐volatile memory and synaptic devices. [ABSTRACT FROM AUTHOR]

Liver tumor semantic segmentation is a crucial task in medical image analysis that requires multiple MRI modalities. This paper proposes a novel coarse-to-fine fusion segmentation approach to detect and segment small liver tumors of various sizes. To enhance the segmentation accuracy of small liver tumors, the method incorporates a detection module and a CSR (convolution-SE-residual) module, which includes a convolution block, an SE (squeeze and excitation) module, and a residual module for fine segmentation. The proposed method demonstrates superior performance compared to conventional single-stage end-to-end networks. A private liver MRI dataset comprising 218 patients with a total of 3605 tumors, including 3273 tumors smaller than 3.0 cm, were collected for the proposed method. There are five types of liver tumors identified in this dataset: hepatocellular carcinoma (HCC); metastases of the liver; cholangiocarcinoma (ICC); hepatic cyst; and liver hemangioma. The results indicate that the proposed method outperforms the single segmentation networks 3D UNet and nnU-Net as well as the fusion networks of 3D UNet and nnU-Net with nnDetection. The proposed architecture was evaluated on a test set of 44 images, with an average Dice similarity coefficient (DSC) and recall of 86.9% and 86.7%, respectively, which is a 1% improvement compared to the comparison method. More importantly, compared to existing methods, our proposed approach demonstrates state-of-the-art performance in segmenting small objects with sizes smaller than 10 mm, achieving a Dice score of 85.3% and a malignancy detection rate of 87.5%. [ABSTRACT FROM AUTHOR]

• A novel physics-informed convolutional neural network f-PICNN for PDEs without any labelled data. • Nonlinear convolutional units (NCUs). • Memory mechanism (numerical results show it can considerably speed up the convergence). • Finite discretization schemes (e.g., Euler, Crank Nicolson and fourth-order Runge-Kutta). • Auto-regressive framework. The physics and interdisciplinary problems in science and engineering are mainly described as partial differential equations (PDEs). Recently, a novel method using physics-informed neural networks (PINNs) to solve PDEs by employing deep neural networks with physical constraints as data-driven models has been pioneered for surrogate modelling and inverse problems. However, the original PINNs based on fully connected neural networks pose intrinsic limitations and poor performance for the PDEs with nonlinearity, drastic gradients, multiscale characteristics or high dimensionality in which the complex features are hard to capture. This leads to difficulties in convergence to correct solutions and high computational costs. To address the above problems, in this paper, a novel physics-informed convolutional neural network framework based on finite discretization schemes with a stack of a series of nonlinear convolutional units (NCUs) for solving PDEs in the space-time domain without any labelled data (f-PICNN) is proposed, in which the memory mechanism can considerably speed up the convergence. Specifically, the initial conditions (ICs) are hard-encoded into the network as the first time-step solution and used to extrapolate the next time-step solution. The Dirichlet boundary conditions (BCs) are constrained by soft BC enforcement while the Neumann BCs are hard enforced. Furthermore, the loss function is designed as a set of discretized PDE residuals and optimized to conform to physics laws. Finally, the proposed auto-regressive model has been proven to be effective in a wide range of 1D and 2D nonlinear PDEs in both space and time under different finite discretization schemes (e.g., Euler, Crank Nicolson and fourth-order Runge-Kutta). The numerical results demonstrate that the proposed framework not only shows the ability to learn the PDEs efficiently but also provides an opportunity for greater conceptual simplicity, and potential for extrapolation from learning the PDEs using a limited dataset. [ABSTRACT FROM AUTHOR]

The advancement of safety and reliability in Nuclear Power Plants (NPP) is essential for ensuring the protection of human life, the environment, and the sustainable use of clean energy. The key factor to achieve this goal is enhancing operator performance, as their decisions play a crucial role during reactor accidents. In this sense, the integration of artificial intelligence (AI) with safety systems in NPP could provide suggestions and recommendations to the operator, leading to a quick and accurate response. However, the complexity of models like Convolutional Neural Networks (CNN), often considered as black boxes, presents challenges for operators to comprehend the decision-making process. This difficulty leads to a lack of trust in the model's output. Therefore, ensuring algorithm transparency becomes crucial to maintain a positive relationship between humans and AI technology. In this study, the necessary data were collected through the automation technique implemented in the (PCTRAN) software. Subsequently, CNN was employed to classify various design-based accidents (DBA), accompanied by transfer learning to optimize accuracy. Finally, Shapley Additive Explanation (SHAP) was used to provide outcome explainability. The results show the effectiveness of automation in reducing time consumption and optimizing hardware resource utilization. Among the pre-trained models, MobileNet exhibits superior performance by requiring a smaller dataset and less time, achieving the highest accuracy, and demonstrating an optimal balance of (correct and incorrect) data prediction within the classes. Furthermore, the use of Explainable AI (XAI) provides result transparency and clarifies the reasons behind CNN incorrect predictions. [ABSTRACT FROM AUTHOR]

Objectives: Convolutional neural network (CNN) is a deep-learning method for image classification and recognition based on a multi-layer NN. In this study, CNN was used to accurately assess cavernous sinus invasion (CSI) in pituitary adenoma (PA). Methods: A total of 371 patients with PA were enrolled in the retrospective study. The cohort was divided into the invasive (n = 102) and non-invasive groups (n = 269) based on surgically confirmed CSI. Images were selected on the T1-enhanced imaging on MR scans. The cohort underwent a fivefold division of randomized datasets for cross-validation. Then, a tenfold augmented dataset (horizontal flip and rotation) of the training set was enrolled in the pre-trained Resnet50 model for transfer learning. The testing set was imported into the trained model for evaluation. Gradient-weighted class activation mapping (Grad-CAM) was used to obtain the occlusion map. The diagnostic values were compared with different dichotomizations of the Knosp grading system (grades 0-1/2-4, 0-2/3a-4, and 0-3a/3b-4). Results: Based on Knosp grades, 20 cases of grade 0, 107 cases of grade 1, 82 cases of grade 2, 104 cases of grade 3a, 22 cases of grade 3b, and 36 cases of grade 4 were recorded. The CSI rates were 0%, 3.7%, 18.3%, 37.5%, 54.5%, and 88.9%. The predicted accuracies of the three dichotomies were 60%, 74%, and 81%. The area under the receiver operating characteristic (AUC-ROC) of Knosp grade for CSI prediction was 0.84; the cutoff was 2.5 with a Youden value of 0.62. The accuracies of the CNN model ranged from 0.80 to 0.96, with AUC-ROC values ranging from 0.89 to 0.98. The Grad-CAM saliency maps confirmed that the region of interest of the model was around the sellar region. Conclusions: We constructed a CNN model with a high proficiency at CSI diagnosis. A more accurate CSI identification was achieved with the constructed CNN than the Knosp grading system. [ABSTRACT FROM AUTHOR]

Chemical mechanical polishing (CMP) is one of the most important and effective technologies to achieve global planarization for precision machining of the wafer surface. The planarization property of slurry is an essential index in evaluating the quality of the copper (Cu) film CMP process. Prediction of the planarization property is of vital significance for product quality control during the CMP process. Data‐driven approaches to predict results based on response surface methodology (RSM) and neural network (NN) to pursue better prediction performance were proposed in this paper. In our design, all three optimization methods (RSM, NN, RSM + NN) are proved to be accurate and reliable, which can predict the experimental results with finer grid accuracy under the condition of only 17 test data points. In particular, the prediction accuracy of RSM + NN method is higher than that of the other two methods, and the error is only 0.16 %. Notably, the results demonstrate that such less time‐consuming optimization methods can realize the acquisition of more desirable CMP process parameters and compositions of slurries, and the versatility and simplicity of our methods can also potentially provide an alternative experiment design concept for a wider range of applications. [ABSTRACT FROM AUTHOR]

The rapid development of neuro-inspired computing demands synaptic devices with ultrafast speed, low power consumption, and multiple non-volatile states, among other features. Here, a high-performance synaptic device is designed and established based on a Ag/PbZr0.52Ti0.48O3 (PZT, (111)-oriented)/Nb:SrTiO3 ferroelectric tunnel junction (FTJ). The advantages of (111)-oriented PZT (~1.2 nm) include its multiple ferroelectric switching dynamics, ultrafine ferroelectric domains, and small coercive voltage. The FTJ shows high-precision (256 states, 8 bits), reproducible (cycle-to-cycle variation, ~2.06%), linear (nonlinearity <1) and symmetric weight updates, with a good endurance of >109 cycles and an ultralow write energy consumption. In particular, manipulations among 150 states are realized under subnanosecond (~630 ps) pulse voltages ≤5 V, and the fastest resistance switching at 300 ps for the FTJs is achieved by voltages <13 V. Based on the experimental performance, the convolutional neural network simulation achieves a high online learning accuracy of ~94.7% for recognizing fashion product images, close to the calculated result of ~95.6% by floating-point-based convolutional neural network software. Interestingly, the FTJ-based neural network is very robust to input image noise, showing potential for practical applications. This work represents an important improvement in FTJs towards building neuro-inspired computing systems. Brain-inspired computing demands high-performance synapses. Here, the authors report a subnanosecond ferroelectric tunnel junction with 256 conductance states, 109 endurance, and 5.3 fJ/bit energy consumption, satisfactory to build synaptic devices. [ABSTRACT FROM AUTHOR]

Background: Accurate segmentation of pelvic bones is an initial step to achieve accurate detection and localisation of pelvic bone metastases. This study presents a deep learning-based approach for automated segmentation of normal pelvic bony structures in multiparametric magnetic resonance imaging (mpMRI) using a 3D convolutional neural network (CNN). Methods: This retrospective study included 264 pelvic mpMRI data obtained between 2018 and 2019. The manual annotations of pelvic bony structures (which included lumbar vertebra, sacrococcyx, ilium, acetabulum, femoral head, femoral neck, ischium, and pubis) on diffusion-weighted imaging (DWI) and apparent diffusion coefficient (ADC) images were used to create reference standards. A 3D U-Net CNN was employed for automatic pelvic bone segmentation. Additionally, 60 mpMRI data from 2020 were included and used to evaluate the model externally. Results: The CNN achieved a high Dice similarity coefficient (DSC) average in both testing (0.80 [DWI images] and 0.85 [ADC images]) and external (0.79 [DWI images] and 0.84 [ADC images]) validation sets. Pelvic bone volumes measured with manual and CNN-predicted segmentations were highly correlated (R2 value of 0.84–0.97) and in close agreement (mean bias of 2.6–4.5 cm3). A SCORE system was designed to qualitatively evaluate the model for which both testing and external validation sets achieved high scores in terms of both qualitative evaluation and concordance between two readers (ICC = 0.904; 95% confidence interval: 0.871–0.929). Conclusions: A deep learning-based method can achieve automated pelvic bone segmentation on DWI and ADC images with suitable quantitative and qualitative performance. [ABSTRACT FROM AUTHOR]

The attention mechanism provides a sequential prediction framework for learning spatial models with enhanced implicit temporal consistency. In this work, we show a systematic design (from 2D to 3D) for how conventional networks and other forms of constraints can be incorporated into the attention framework for learning long-range dependencies for the task of pose estimation. The contribution of this paper is to provide a systematic approach for designing and training of attention-based models for the end-to-end pose estimation, with the flexibility and scalability of arbitrary video sequences as input. We achieve this by adapting temporal receptive field via a multi-scale structure of dilated convolutions. Besides, the proposed architecture can be easily adapted to a causal model enabling real-time performance. Any off-the-shelf 2D pose estimation systems, e.g. Our method achieves the state-of-the-art performance and outperforms existing methods by reducing the mean per joint position error to 33.4mm on Human 3.6M dataset. Our code is available at https://github.com/lrxjason/Attention3DHumanPose [ABSTRACT FROM AUTHOR]

Background and objective. Early detection of hepatocellular carcinoma (HCC) is crucial for clinical management. Current studies have reported large HCC detections using automatic algorithms, but there is a lack of research on automatic detection of small HCCs (sHCCs). This study is to investigate the feasibility of automatic detection of sHCC (≤2 cm) based on pattern matching and deep learning (PM-DL) model. Materials and methods. A retrospective study included 5376 image sets from 56 cirrhosis patients (28 sHCC patients with 32 pathologically confirmed lesions and 28 non-HCC cirrhosis patients) in the training-validation cohort to build and validate the model through five-fold cross-validation. In addition, an external test cohort including 6144 image sets from 64 cirrhosis patients (32 sHCC patients with 38 lesions and 32 non-HCC cirrhosis patients) was applied to further verify the generalization ability of the model. The proposed PM-DL model consisted of three main steps: 3D co-registration and liver segmentation, screening of suspicious lesions on diffusion-weighted imaging images based on pattern matching algorithm, and identification/segmentation of sHCC lesions on dynamic contrast-enhanced images with convolutional neural network. Results. The PM-DL model achieved a sensitivity of 89.74% and a positive predictive value of 85.00% in the external test cohort for per-lesion analysis. No significant difference was observed in volumes (P = 0.13) and the largest sizes (P = 0.89) between manually delineated and segmented lesions. The DICE coefficient reached 0.77 ± 0.16. Similar performances were identified in the validation cohort. Moreover, the PM-DL model outperformed Liver Imaging Reporting and Data System (LI-RADS) in sensitivity (probable HCCs: LR-5 or LR-4, P = 0.18; definite HCCs: LR-5, P < 0.001), with a similar high specificity for per-patient analysis. Conclusion. The PM-DL model may be feasible for accurate automatic detection of sHCC in cirrhotic liver. [ABSTRACT FROM AUTHOR]

Image inpainting algorithms have a wide range of applications, which can be used for object removal in digital images. With the development of semantic level image inpainting technology, this brings great challenges to blind image forensics. In this case, many conventional methods have been proposed which have disadvantages such as high time complexity and low robustness to postprocessing operations. Therefore, this paper proposes a mask regional convolutional neural network (Mask R-CNN) approach for patch-based inpainting detection. According to the current research, many deep learning methods have shown the capacity for segmentation tasks when labeled datasets are available, so we apply a deep neural network to the domain of inpainting forensics. This deep learning model can distinguish and obtain different features between the inpainted and noninpainted regions. To reduce the missed detection areas and improve detection accuracy, we also adjust the sizes of the anchor scales due to the inpainting images and replace the original nonmaximum suppression single threshold with an improved nonmaximum suppression (NMS). The experimental results demonstrate this intelligent method has better detection performance over recent approaches of image inpainting forensics. [ABSTRACT FROM AUTHOR]

Synthetic aperture radar (SAR) is a powerful tool for monitoring sea states in terms of the significant wave height (SWH). Regarding the specific wave mode, to date, the previous empirical models for estimating SWH from SAR data rely on single polarization. In the emerging deep learning era, few published quad-polarized SAR SWH retrieval algorithms have been based on machine learning technique, and whether quad-polarimetry improves the skill of wave height estimation remains a question. Here we propose a deep residual convolutional neural network-based SAR SWH retrieval algorithm in quad-polarization. By collocating WaveWatch III sea state hindcasts and all available archives of quad-polarized Chinese Gaofen-3 SAR imagettes in wave mode, a database with approximately 30,000 matchups was employed to establish our deeply-learned network. The GaoFen-3 significant wave height retrievals were validated against the hindcast dataset independent of training along with altimeter observations. The result of good consistency in terms of a root mean square error of 0.32 m (under sea state conditions of approximately 0.5–7.0 m) outperforms the existing Gaofen-3 wave height retrieval algorithms. Additionally, this paper introduces a discussion about the contribution of polarizations by comparing SWH derived from single-, dual- and quad-polarized deep convolutional neural networks. Single-polarized Gaofen-3 SAR data are found to be sufficient to provide accurate estimates compared to quad-polarization via a deep learning model under moderate sea conditions. Exploitation of SAR quad-polarimetry information will improve SAR wave height retrievals under high sea conditions. • We propose residual networks for estimating wave heights from Gaofen-3. • Our approach outperformed the existing Gaofen-3 wave height retrieval methods. • Single polarization is sufficient for accurately estimating moderate wave heights. • Quad-polarized data enhance the SAR wave estimating capabilities for high states. [ABSTRACT FROM AUTHOR]

• This work introduced a framework combining the CNN and LSTM networks to extract time-frequency and temporal dependence features from the Korotkoff sound signal for automatic non-invasive blood pressure measurement. • The proposed deep learning based automatic auscultatory method achieved accurate blood pressure measurement on subjects with a wide blood pressure range. • The proposed method also achieved better blood pressure category classification performance in comparison with the traditional oscillometric method. Manual auscultatory is the gold standard for clinical non-invasive blood pressure (BP) measurement, but its usage is decreasing as it requires substantial professional skills and training, and its environmental concerns related to mercury toxicity. As an alternative, automatic oscillometric technique has been used as one of the most common methods for BP measurement, however, it only estimates BPs based on empirical equations. To overcome these problems, this study aimed to develop a deep learning-based automatic auscultatory BP measurement method, and clinically validate its performance. A deep learning-based method that utilized time-frequency characteristics and temporal dependence of segmented Korotkoff sound (KorS) signals and employed convolutional neural network (CNN) and long short-term memory (LSTM) network was developed and trained using KorS and cuff pressure signals recorded from 314 subjects. The BPs determined by the manual auscultatory method was used as the reference for each measurement. The measurement error and BP category classification performance of our proposed method were then validated on a separate dataset of 114 subjects. Its performance in comparison with the oscillometric method was also comprehensively analyzed. The deep learning method achieved measurement errors of 0.2 ± 4.6 mmHg and 0.1 ± 3.2 mmHg for systolic BP and diastolic BP, respectively, and achieved high sensitivity, specificity and accuracy (all > 90 %) in classifying hypertensive subjects, which were better than those of the traditional oscillometric method. This validation study demonstrated that deep learning-based automatic auscultatory BP measurement can be developed to achieve high measurement accuracy and high BP category classification performance. [ABSTRACT FROM AUTHOR]

Quantitative susceptibility mapping (QSM) has demonstrated great potential in quantifying tissue susceptibility in various brain diseases. However, the intrinsic ill-posed inverse problem relating the tissue phase to the underlying susceptibility distribution affects the accuracy for quantifying tissue susceptibility. Recently, deep learning has shown promising results to improve accuracy by reducing the streaking artifacts. However, there exists a mismatch between the observed phase and the theoretical forward phase estimated by the susceptibility label. In this study, we proposed a model-based deep learning architecture that followed the STI (susceptibility tensor imaging) physical model, referred to as MoDL-QSM. Specifically, MoDL-QSM accounts for the relationship between STI-derived phase contrast induced by the susceptibility tensor terms (χ 13 , χ 23 and χ 33) and the acquired single-orientation phase. The convolutional neural networks are embedded into the physical model to learn a regularization term containing prior information. χ 33 and phase induced by χ 13 and χ 23 terms were used as the labels for network training. Quantitative evaluation metrics were compared with recently developed deep learning QSM methods. The results showed that MoDL-QSM achieved superior performance, demonstrating its potential for future applications. [Display omitted] [ABSTRACT FROM AUTHOR]

Deformable image registration (DIR) between daily and reference images is fundamentally important for adaptive radiotherapy. In the last decade, deep learning-based image registration methods have been developed with faster computation time and improved robustness compared to traditional methods. However, the registration performance is often degraded in extra-cranial sites with large volume containing multiple anatomic regions, such as Computed Tomography (CT)/Magnetic Resonance (MR) images used in head and neck (HN) radiotherapy. In this study, we developed a hierarchical deformable image registration (DIR) framework, Patch-based Registration Network (Patch-RegNet), to improve the accuracy and speed of CT-MR and MR-MR registration for head-and-neck MR-Linac treatments. Patch-RegNet includes three steps: a whole volume global registration, a patch-based local registration, and a patch-based deformable registration. Following a whole-volume rigid registration, the input images were divided into overlapping patches. Then a patch-based rigid registration was applied to achieve accurate local alignment for subsequent DIR. We developed a ViT-Morph model, a combination of a convolutional neural network (CNN) and the Vision Transformer (ViT), for the patch-based DIR. A modality independent neighborhood descriptor was adopted in our model as the similarity metric to account for both inter-modality and intra-modality registration. The CT-MR and MR-MR DIR models were trained with 242 CT-MR and 213 MR-MR image pairs from 36 patients, respectively, and both tested with 24 image pairs (CT-MR and MR-MR) from 6 other patients. The registration performance was evaluated with 7 manually contoured organs (brainstem, spinal cord, mandible, left/right parotids, left/right submandibular glands) by comparing with the traditional registration methods in Monaco treatment planning system and the popular deep learning-based DIR framework, Voxelmorph. Evaluation results show that our method outperformed VoxelMorph by 6 % for CT-MR registration, and 4 % for MR-MR registration based on DSC measurements. Our hierarchical registration framework has been demonstrated achieving significantly improved DIR accuracy of both CT-MR and MR-MR registration for head-and-neck MR-guided adaptive radiotherapy. • Enable rapid and accurate inter- and intra-modality deformable registration. • A hierarchical registration framework using patch-based registration. • Taking advantage of both convolutional neural network and transformer. • Benefit the online MR-guided adaptive radiotherapy for head and neck cancer. [ABSTRACT FROM AUTHOR]

• image-to-image regression for surrogate modelling for groundwater flow prediction, • use of attention mechanism to identify and focus the learning on salient regions of the domain, • significantly reduced computational time compared to a state-of-the-art numerical solver. Numerical simulations of groundwater flow are used to analyze and predict the response of an aquifer system to its change in state by approximating the solution of the fundamental groundwater physical equations. The most used and classical methodologies, such as Finite Difference (FD) and Finite Element (FE) Methods, use iterative solvers which are associated with high computational cost. This study proposes a physics-based convolutional encoder-decoder neural network as a surrogate model to quickly calculate the response of the groundwater system. Holding strong promise in cross-domain mappings, encoder-decoder networks are applicable for learning complex input-output mappings of physical systems. This manuscript presents an Attention U-Net model that attempts to capture the fundamental input-output relations of the groundwater system and generates solutions of hydraulic head in the whole domain given a set of physical parameters and boundary conditions. The model accurately predicts the steady state response of a highly heterogeneous groundwater system given the locations and piezometric head of up to 3 wells as input. The network learns to pay attention only in the relevant parts of the domain and the generated hydraulic head field corresponds to the target samples in great detail. Even relative to coarse finite difference approximations the proposed model is shown to be significantly faster than a comparative state-of-the-art numerical solver, thus providing a base for further development of the presented networks as surrogate models for groundwater prediction. [ABSTRACT FROM AUTHOR]

• A deep learning-based phase reconstruction scheme is proposed for high-resolution multi-shot (MSH) DWI reconstruction. • Higher SNR (especially with high acceleration rates), fewer aliasing artifacts, lower ghost-to-signal-ratio (GSR), higher tracked fiber counts, and finer fiber delineation can be obtained with MSH-DWI reconstruction compared to conventional MUSE. • Single-shot (SSH) DWIs were used for training, making the proposed method readily applicable for routine clinical exams. A phase correction method for high-resolution multi-shot (MSH) diffusion weighted imaging (DWI) is proposed. The efficacy and generalization capability of the method were validated on both healthy volunteers and patients. Conventionally, inter-shot phase variations for MSH echo-planar imaging (EPI) DWI are corrected by model-based algorithms. However, many acquisition imperfections are hard to measure accurately for conventional model-based methods, making the phase estimation and artifacts suppression unreliable. We propose a deep learning multiplexed sensitivity-encoding (DL-MUSE) framework to improve the phase estimations based on convolutional neural network (CNN) reconstruction. Aliasing-free single-shot (SSH) DW images, which have been used routinely in clinical settings, were used for training before the aliasing correction of MSH-DWI images. A dual-channel U-net comprising multiple convolutional layers was used for the phase estimation of MSH-EPI. The network was trained on a dataset containing 30 healthy volunteers and tested on another dataset of 52 healthy subjects and 15 patients with lesions or tumors with different shot numbers (4, 6 and 8). To further validate the generalization capability of our network, we acquired a dataset with different numbers of shots, TEs, partial Fourier factors, resolutions, ETLs, FOVs, coil numbers, and image orientations from two sites. We also compared the reconstruction performance of our proposed method with that of the conventional MUSE and SSH-EPI qualitatively and quantitatively. Our results show that DL-MUSE is capable of correcting inter-shot phase errors with high and robust performance. Compared to conventional model-based MUSE, our method, by applying deep learning-based phase corrections, showed reduced distortion, noise level, and signal loss in high b -value DWIs. The improvements of image quality become more evident as the shot number increases from 4 to 8, especially in those central regions of the images, where g -factor artifacts are severe. Furthermore, the proposed method could provide the information about the orientation of the white matter with better consistency and achieve finer fibers delineation compared to the SSH-EPI method. Besides, the experiments on volunteers and patients from two different sites demonstrated the generalizability of our proposed method preliminarily. A deep learning-based reconstruction algorithm for MSH-EPI images, which helps improve image quality greatly, was proposed. Results from healthy volunteers and tumor patients demonstrated the feasibility and generalization performances of our method for high-resolution MSH-EPI DWI, which can be used for routine clinical applications as well as neuroimaging research. [ABSTRACT FROM AUTHOR]