Sleep quality is heavily influenced by sleep posture, with research indicating that a supine posture can worsen obstructive sleep apnea (OSA) while lateral postures promote better sleep. For patients confined to beds, regular changes in posture are crucial to prevent the development of ulcers and bedsores. This study presents a novel sparse sensor-based spatiotemporal convolutional neural network (S³CNN) for detecting sleep posture. This S³CNN holistically incorporates a pair of spatial convolution neural networks to capture cardiorespiratory activity maps and a pair of temporal convolution neural networks to capture the heart rate and respiratory rate. Sleep data were collected in actual sleep conditions from 22 subjects using a sparse sensor array. The S³CNN was then trained to capture the spatial pressure distribution from the cardiorespiratory activity and temporal cardiopulmonary variability from the heart and respiratory data. Its performance was evaluated using three rounds of 10 fold cross-validation on the 8583 data samples collected from the subjects. The results yielded 91.96% recall, 92.65% precision, and 93.02% accuracy, which are comparable to the state-of-the-art methods that use significantly more sensors for marginally enhanced accuracy. Hence, the proposed S³CNN shows promise for sleep posture monitoring using sparse sensors, demonstrating potential for a more cost-effective approach.

Deep learning has achieved remarkable progress, particularly in neuroimaging analysis. Deep learning applications have also been extended from adult to pediatric medical images, and thus, this paper aims to present a systematic review of this recent research. We first introduce the commonly used deep learning methods and architectures in neuroimaging, such as convolutional neural networks, auto-encoders, and generative adversarial networks. A non-exhaustive list of commonly used publicly available pediatric neuroimaging datasets and repositories are included, followed by a categorical review of recent works in pediatric MRI-based deep learning studies in the past five years. These works are categorized into recognizing neurodevelopmental disorders, identifying brain and tissue structures, estimating brain age/maturity, predicting neurodevelopment outcomes, and optimizing MRI brain imaging and analysis. Finally, we also discuss the recent achievements and challenges on these applications of deep learning to pediatric neuroimaging.

Objective: To explore the value of machine learning models in preoperative prediction of microvascular invasion (MVI) in hepatocellular carcinoma (HCC) based on dual-phase contrast-enhanced CT radiomics features. Methods: The data of 148 patients [106 males and 42 females, with an average age of (58±11) years] with HCC confirmed by pathology in the First Affiliated Hospital of Soochow University from January 2015 to May 2020 were retrospectively analyzed, including 88 cases of positive MVI and 60 cases of negative MVI. According to the ratio of 7∶3, the patients were randomly divided into the training and validation sets, respectively. The three-dimensional (3D) radiomics features of HCC in arterial phase (AP) and portal venous phase (PP) were extracted by MaZda software, and the optimal feature subset was obtained by combining three feature selection methods (FPM method) and Lasso regression. Then, six machine learning methods were used to build the prediction models. Receiver operating characteristic (ROC) curves were drawn to evaluate the prediction ability of the aforementioned models, and the area under the curve (AUC), accuracy, sensitivity and specificity were calculated. Results: Radiomics features of HCC in AP and PP were extracted by MaZda software, with 239 in each phase. There were 7 optimal features in AP and 14 optimal features in PP selected by FPM method and Lasso regression, respectively. The AUCs of decision tree, extreme gradient boosting, random forest, support vector machine (SVM), generalized linear model, and neural network based on the 7 optimal features in AP in the validation set were 0.736, 0.910, 0.913, 0.915, 0.897, 0.648, respectively. The SVM had the highest AUC in the validation set, with the accuracy, sensitivity and specificity of 95.35%, 95.83% and 94.74%, respectively. Likewise, the AUCs of machine learning models in prediction of MVI in HCC based on the 14 optimal features in PP in the validation set were 0.873, 0.876, 0.913, 0.859, 0.877, 0.834, respectively, and there were no significant differences (all P>0.05). The random forest had the highest AUC in the validation set, with the accuracy, sensitivity and specificity of 90.70%, 87.50% and 94.74%, respectively. Conclusion: Machine learning models based on dual-phase enhanced CT radiomics features can be used in preoperative prediction of MVI in HCC, particularly the SVM and random forest models have high prediction efficiency.

Rapid identification of large vessel occlusion (LVO) is essential for timely thrombectomy; yet, definitive diagnosis requires vascular imaging that is unavailable in many prehospital settings. We investigated whether resting-state functional near-infrared spectroscopy (fNIRS) captures spatially organized signals associated with LVO and can support task-free classification within a passive brain–computer interface framework. Seventy-seven patients with suspected stroke were enrolled, and 47 met prespecified quality-control and diagnostic criteria (17 LVO and 30 non-LVO ischemic stroke). We compared channel-level and anatomically grouped regional features using age- and sex-adjusted multivariate analyses and participant-level cross-validation. Channel-level differences did not remain significant after adjustment, whereas regional multivariate effects persisted, particularly in sensorimotor regions. The regional time-domain model achieved 82.4% accuracy, 76.2% sensitivity, 86.0% specificity, and an area under the receiver operating characteristic curve of 0.832 in internal fixed-configuration cross-validation. Predictive contributions were distributed across all four predefined regions, while adding feature families through early or late fusion did not improve performance. These convergent statistical and classification results indicate that LVO-associated fNIRS information is spatially distributed and is captured more effectively by anatomical aggregation than by isolated-channel analysis or greater feature complexity. The findings support multicenter evaluation of task-free fNIRS for early LVO risk assessment.