Segmenting the left ventricle from the transgastric short-axis views (TSVs) on transesophageal echocardiography (TEE) is the cornerstone for cardiovascular assessment during perioperative management. Even for seasoned professionals, the procedure remains time-consuming and experience-dependent. The current study aims to evaluate the feasibility of deep learning for automatic segmentation by assessing the validity of different U-Net algorithms. A large dataset containing 1388 TSV acquisitions was retrospectively collected from 451 patients (32% women, average age 53.42 years) who underwent perioperative TEE between July 2015 and October 2023. With image preprocessing and data augmentation, 3336 images were included in the training set, 138 images in the validation set, and 138 images in the test set. Four deep neural networks (U-Net, Attention U-Net, UNet++, and UNeXt) were employed for left ventricle segmentation and compared in terms of the Jaccard similarity coefficient (JSC) and Dice similarity coefficient (DSC) on the test set, as well as the number of network parameters, training time, and inference time. The Attention U-Net and U-Net++ models performed better in terms of JSC (the highest average JSC: 86.02%) and DSC (the highest average DSC: 92.00%), the UNeXt model had the smallest network parameters (1.47 million), and the U-Net model had the least training time (6428.65 s) and inference time for a single image (101.75 ms). The Attention U-Net model outperformed the other three models in challenging cases, including the impaired boundary of left ventricle and the artifact of the papillary muscle. This pioneering exploration demonstrated the feasibility of deep learning for the segmentation of the left ventricle from TSV on TEE, which will facilitate an accelerated and objective alternative of cardiovascular assessment for perioperative management.

The vulnerability of modern neural networks to random noise and deliberate attacks has raised concerns about their robustness, particularly as they are increasingly utilized in safety- and security-critical applications. Although recent research efforts were made to enhance robustness through retraining with adversarial examples or employing data augmentation techniques, a comprehensive investigation into the effects of training data perturbations on model robustness remains lacking. This paper presents the first extensive empirical study investigating the influence of data perturbations during model retraining. The experimental analysis focuses on both random and adversarial robustness, following established practices in the field of robustness analysis. Various types of perturbations in different aspects of the dataset are explored, including input, label, and sampling distribution. Single-factor and multi-factor experiments are conducted to assess individual perturbations and their combinations. The findings provide insights into constructing high-quality training datasets for optimizing robustness and recommend the appropriate degree of training set perturbations that balance robustness and correctness, and contribute to understanding model robustness in deep learning and offer practical guidance for enhancing model performance through perturbed retraining, promoting the development of more reliable and trustworthy deep learning systems for safety-critical applications.

OBJECTIVE: In pediatric medicine, precise estimation of bone age is essential for skeletal maturity evaluation, growth disorder diagnosis, and therapeutic intervention planning. Conventional techniques for determining bone age depend on radiologists\' subjective judgments, which may lead to non-negligible differences in the estimated bone age. This study proposes a deep learning-based model utilizing a fully connected convolutional neural network(CNN) to predict bone age from left-hand radiographs.
METHODS: The data set used in this study, consisting of 473 patients, was retrospectively retrieved from the PACS (Picture Achieving and Communication System) of a single institution. We developed a fully connected CNN consisting of four convolutional blocks, three fully connected layers, and a single neuron as output. The model was trained and validated on 80% of the data using the mean-squared error as a cost function to minimize the difference between the predicted and reference bone age values through the Adam optimization algorithm. Data augmentation was applied to the training and validation sets yielded in doubling the data samples. The performance of the trained model was evaluated on a test data set (20%) using various metrics including, the mean absolute error (MAE), median absolute error (MedAE), root-mean-squared error (RMSE), and mean absolute percentage error (MAPE). The code of the developed model for predicting the bone age in this study is available publicly on GitHub at https://github.com/afiosman/deep-learning-based-bone-age-estimation .
RESULTS: Experimental results demonstrate the sound capabilities of our model in predicting the bone age on the left-hand radiographs as in the majority of the cases, the predicted bone ages and reference bone ages are nearly close to each other with a calculated MAE of 2.3 [1.9, 2.7; 0.95 confidence level] years, MedAE of 2.1 years, RMAE of 3.0 [1.5, 4.5; 0.95 confidence level] years, and MAPE of 0.29 (29%) on the test data set.
CONCLUSIONS: These findings highlight the usability of estimating the bone age from left-hand radiographs, helping radiologists to verify their own results considering the margin of error on the model. The performance of our proposed model could be improved with additional refining and validation.

OBJECTIVE: The Alberta Stroke Program Early CT Score (ASPECTS), a systematic method for assessing ischemic changes in acute ischemic stroke using non-contrast computed tomography (NCCT), is often interpreted relying on expert experience and can vary between readers. This study aimed to develop a clinically applicable automatic ASPECTS system employing deep learning (DL).
METHODS: This study enrolled 1987 NCCT scans that were retrospectively collected from four centers between January 2017 and October 2021. A DL-based system for automated ASPECTS assessment was trained on a development cohort (N = 1767) and validated on an independent test cohort (N = 220). The consensus of experienced physicians was regarded as a reference standard. The validity and reliability of the proposed system were assessed against physicians\' readings. A real-world prospective application study with 13,399 patients was used for system validation in clinical contexts.
RESULTS: The DL-based system achieved an area under the receiver operating characteristic curve (AUC) of 84.97% and an intraclass correlation coefficient (ICC) of 0.84 for overall-level analysis on the test cohort. The system\'s diagnostic sensitivity was 94.61% for patients with dichotomized ASPECTS at a threshold of ≥ 6, with substantial agreement (ICC = 0.65) with expert ratings. Combining the system with physicians improved AUC from 67.43 to 89.76%, reducing diagnosis time from 130.6 ± 66.3 s to 33.3 ± 8.3 s (p < 0.001). During the application in clinical contexts, 94.0% (12,591) of scans successfully processed by the system were utilized by clinicians, and 96% of physicians acknowledged significant improvement in work efficiency.
CONCLUSIONS: The proposed DL-based system could accurately and rapidly determine ASPECTS, which might facilitate clinical workflow for early intervention.
CONCLUSIONS: The deep learning-based automated ASPECTS evaluation system can accurately and rapidly determine ASPECTS for early intervention in clinical workflows, reducing processing time for physicians by 74.8%, but still requires validation by physicians when in clinical applications.
CONCLUSIONS: The deep learning-based system for ASPECTS quantification has been shown to be non-inferior to expert-rated ASPECTS. This system improved the consistency of ASPECTS evaluation and reduced processing time to 33.3 seconds per scan. 94.0% of scans successfully processed by the system were utilized by clinicians during the prospective clinical application.

Determining sources and spatial distributions of potentially toxic elements (PTEs) is a crucial issue of soil pollution survey. However, uncertainty estimation for source contributions remains lack, and accurate spatial prediction is still challenging. Robust Bayesian multivariate receptor model (RBMRM) was applied to the soil dataset of Qingzhou City (8 PTEs in 429 samples), to calculate source contributions with uncertainties. Multi-task convolutional neural network (MTCNN) was proposed to predict spatial distributions of soil PTEs. RBMRM afforded three sources, consistent with US-EPA positive matrix factorization. Natural source dominated As, Cr, Cu, and Ni contents (78.5 %∼86.1 %), and contributed 37.1 %, 61.0 %, and 65.9 % of Cd, Pb, and Zn, exhibiting low uncertainties with uncertainty index (UI) < 26.7 %. Industrial, traffic, and agricultural sources had significant influences on Cd, Pb, and Zn (30.2 %∼61.9 %), with UI < 39.3 %. Hg originated dominantly from atmosphere deposition (99.1 %), with relatively high uncertainties (UI=87.7 %). MTCNN acquired satisfactory accuracies, with R2 of 0.357-0.896 and nRMSE of 0.092-0.366. Spatial distributions of As, Cd, Cr, Cu, Ni, Pb, and Zn were influenced by parent materials. Cd, Hg, Pb, and Zn showed significant hotspot in urban area. This work conducted a new approach exploration, and practical implications for soil pollution regulation were proposed.

OBJECTIVE: Laparoscopic distal gastrectomy (LDG) is a difficult procedure for early career surgeons. Artificial intelligence (AI)-based surgical step recognition is crucial for establishing context-aware computer-aided surgery systems. In this study, we aimed to develop an automatic recognition model for LDG using AI and evaluate its performance.
METHODS: Patients who underwent LDG at our institution in 2019 were included in this study. Surgical video data were classified into the following nine steps: (1) Port insertion; (2) Lymphadenectomy on the left side of the greater curvature; (3) Lymphadenectomy on the right side of the greater curvature; (4) Division of the duodenum; (5) Lymphadenectomy of the suprapancreatic area; (6) Lymphadenectomy on the lesser curvature; (7) Division of the stomach; (8) Reconstruction; and (9) From reconstruction to completion of surgery. Two gastric surgeons manually assigned all annotation labels. Convolutional neural network (CNN)-based image classification was further employed to identify surgical steps.
RESULTS: The dataset comprised 40 LDG videos. Over 1,000,000 frames with annotated labels of the LDG steps were used to train the deep-learning model, with 30 and 10 surgical videos for training and validation, respectively. The classification accuracies of the developed models were precision, 0.88; recall, 0.87; F1 score, 0.88; and overall accuracy, 0.89. The inference speed of the proposed model was 32 ps.
CONCLUSIONS: The developed CNN model automatically recognized the LDG surgical process with relatively high accuracy. Adding more data to this model could provide a fundamental technology that could be used in the development of future surgical instruments.

Closed-loop neurofeedback training utilizes neural signals such as scalp electroencephalograms (EEG) to manipulate specific neural activities and the associated behavioral performance. A spatiotemporal filter for high-density whole-head scalp EEG using a convolutional neural network can overcome the ambiguity of the signaling source because each EEG signal includes information on the remote regions. We simultaneously acquired EEG and functional magnetic resonance images in humans during the brain-computer interface (BCI) based neurofeedback training and compared the reconstructed and modeled hemodynamic responses of the sensorimotor network. Filters constructed with a convolutional neural network captured activities in the targeted network with spatial precision and specificity superior to those of the EEG signals preprocessed with standard pipelines used in BCI-based neurofeedback paradigms. The middle layers of the trained model were examined to characterize the neuronal oscillatory features that contributed to the reconstruction. Analysis of the layers for spatial convolution revealed the contribution of distributed cortical circuitries to reconstruction, including the frontoparietal and sensorimotor areas, and those of temporal convolution layers that successfully reconstructed the hemodynamic response function. Employing a spatiotemporal filter and leveraging the electrophysiological signatures of the sensorimotor excitability identified in our middle layer analysis would contribute to the development of a further effective neurofeedback intervention.

UNASSIGNED: The increasing adoption of intestinal ultrasound (IUS) for monitoring inflammatory bowel diseases (IBD) by IBD providers has uncovered new challenges regarding standardized image interpretation and limitations as a research tool. Artificial intelligence approaches can help address these challenges. We aim to determine the feasibility of radiomic analysis of IUS images and to determine if a radiomics-based classification model can accurately differentiate between normal and abnormal IUS images. We will also compare the radiomic-based model\'s performance to a convolutional neural network (CNN)-based classification model to understand which method is more effective for extracting meaningful information from IUS images.
UNASSIGNED: Retrospectively analyzing IUS images obtained during routine outpatient visits, we developed and tested radiomic-based and CNN-based models to distinguish between normal and abnormal images, with abnormal images defined as bowel wall thickness > 3 mm or bowel hyperemia with modified Limberg score ≥ 1 (both are surrogate markers for inflammation). Model performances were measured by area under the receiver operator curve (AUC).
UNASSIGNED: For this feasibility study, 125 images (33% abnormal) were analyzed. A radiomic-based model using XG boost yielded the best classifier model with average test AUC 0.98%, 93.8% sensitivity, 93.8% specificity, and 93.7% accuracy. The CNN-based classification model yielded an average testing AUC of 0.75.
UNASSIGNED: Radiomic analysis of IUS images is feasible, and a radiomic-based classification model could accurately differentiate abnormal from normal images. Our findings establish methods to facilitate future radiomic-based IUS studies that can help standardize image interpretation and expand IUS research capabilities.

This study introduces a deep-learning-based automatic sleep scoring system to detect sleep apnea using a single-lead electrocardiography (ECG) signal, focusing on accurately estimating the apnea-hypopnea index (AHI). Unlike other research, this work emphasizes AHI estimation, crucial for the diagnosis and severity evaluation of sleep apnea. The suggested model, trained on 1465 ECG recordings, combines the deep-shallow fusion network for sleep apnea detection network (DSF-SANet) and gated recurrent units (GRUs) to analyze ECG signals at 1-min intervals, capturing sleep-related respiratory disturbances. Achieving a 0.87 correlation coefficient with actual AHI values, an accuracy of 0.82, an F1 score of 0.71, and an area under the receiver operating characteristic curve of 0.88 for per-segment classification, our model was effective in identifying sleep-breathing events and estimating the AHI, offering a promising tool for medical professionals.

BACKGROUND: Anterior cruciate ligament (ACL) injuries are common in sports and are critical knee injuries that require prompt diagnosis. Magnetic resonance imaging (MRI) is a strong, noninvasive tool for detecting ACL tears, which requires training to read accurately. Clinicians with different experiences in reading MR images require different information for the diagnosis of ACL tears. Artificial intelligence (AI) image processing could be a promising approach in the diagnosis of ACL tears.
OBJECTIVE: This study sought to use AI to (1) diagnose ACL tears from complete MR images, (2) identify torn-ACL images from complete MR images with a diagnosis of ACL tears, and (3) differentiate intact-ACL and torn-ACL MR images from the selected MR images.
METHODS: The sagittal MR images of torn ACL (n=1205) and intact ACL (n=1018) from 800 cases and the complete knee MR images of 200 cases (100 torn ACL and 100 intact ACL) from patients aged 20-40 years were retrospectively collected. An AI approach using a convolutional neural network was applied to build models for the objective. The MR images of 200 independent cases (100 torn ACL and 100 intact ACL) were used as the test set for the models. The MR images of 40 randomly selected cases from the test set were used to compare the reading accuracy of ACL tears between the trained model and clinicians with different levels of experience.
RESULTS: The first model differentiated between torn-ACL, intact-ACL, and other images from complete MR images with an accuracy of 0.9946, and the sensitivity, specificity, precision, and F1-score were 0.9344, 0.9743, 0.8659, and 0.8980, respectively. The final accuracy for ACL-tear diagnosis was 0.96. The model showed a significantly higher reading accuracy than less experienced clinicians. The second model identified torn-ACL images from complete MR images with a diagnosis of ACL tear with an accuracy of 0.9943, and the sensitivity, specificity, precision, and F1-score were 0.9154, 0.9660, 0.8167, and 0.8632, respectively. The third model differentiated torn- and intact-ACL images with an accuracy of 0.9691, and the sensitivity, specificity, precision, and F1-score were 0.9827, 0.9519, 0.9632, and 0.9728, respectively.
CONCLUSIONS: This study demonstrates the feasibility of using an AI approach to provide information to clinicians who need different information from MRI to diagnose ACL tears.