Sleep is essential for human survival, and accurate sleep staging has significant potential for real-world applications. While deep learning methods for automatic sleep staging have shown promise, several challenges remain: 1) how to efficiently extract multimodal representations with strong generalization; 2) how to leverage the similarities and differences in multimodal signals for accurate sleep staging; and 3) how to utilize abundant unlabeled data to enhance model practicality. To address these challenges, we propose MultiConsSleepNet, a multimodal consistency-based sleep staging network. It integrates a unimodal feature extractor and a multimodal consistency extractor to capture local-domain representations of electroencephalogram (EEG) and electrooculogram (EOG) signals, enhancing feature consistency both within and across modalities. We also incorporate self-supervised contrastive learning strategies for both unimodal and multimodal consistency learning, thereby improving model generalization and transferability. Experiments on three datasets demonstrate that MultiConsSleepNet achieves state-of-the-art performance with limited labeled data, effectively leveraging unlabeled data to enhance application value.Clinical relevance— This study highlights the potential of a multimodal deep learning approach for automated sleep staging, achieving high performance with only 10% labeled data. Specifically, it attains an average F1 score of 74.2% and accuracy of 79.6% in cross-dataset scenarios. This method provides a valuable tool for clinicians, enabling efficient and reliable sleep staging with minimal labeled data. It can improve the speed and consistency of sleep disorder diagnoses while reducing reliance on manual analysis, thereby enhancing clinical workflow efficiency.
Sleep staging is an essential step in sleep quality assessment and sleep disorder diagnosis. However, most neural network approaches to automatic sleep staging have focused on improving network architectures and data augmentation techniques, the latent potential of sleep staging labels has not been fully explored. Existing approaches often use the one-hot encoding method, which encodes each sleep stage as an independent and parallel label, fail to capture the prior information and relationships among sleep stages, thereby limiting their potential. In this study, we propose a novel label encoding method called Binary Carry-Forward Encoding (BCFE) for sleep staging classification. Unlike traditional approaches with one-hot encoding, we first decompose the 5-class sleep staging task into four binary classification tasks aligned with the logical progression of sleep stages. By introducing 4-dimensional binary vectors, BCFE embeds potentially prior information into the label structure. Predicted labels are generated through Human-Like Progression (HLP) inference, and training is optimized with a Masked Binary CrossEntropy (MBCE) loss. These methods enables the network to better utilize nuanced label characteristics, enhancing classification performance. The evaluation was conducted across combinations of three models (DeepSleepNet, AttnSleep and SleePyCo) and two datasets (ISRUC-S3 and Sleep-EDF). The results showed that our proposed method demonstrated significant improvements compared to the traditional One-Hot + Cross-Entropy method, particularly for the DeepSleepNet architecture on the ISRUC-S3 dataset. Notably, it achieved nearly 7% increase in performance (Accuracy: 73.7% → 80.3%; Macro-F1: 71.3% → 78.8%; kappa: 66.7% → 74.6%).
Shape registration between patient-specific organ geometries and endoscopic camera images is crucial for image-guided surgery. Although machine learning methods have been investigated for this task, collecting training datasets remains challenging because of the limited use of three-dimensional (3D) imaging in surgical settings. Offline learning with synthetic images generated from preoperative 3D-CT data has been proposed; however, ensuring robust learning under domain discrepancies between synthetic and real images remains a major challenge. In this study, we propose a diffusion-based offline learning strategy to register the shape a liver mesh in laparoscopic camera images. Within this framework, semantic organ labels serve as an image feature shared between synthetic and real-world images, thereby mitigating the domain gap and facilitating accurate registration. During model training, Gaussian noise is introduced into the registration parameters, and the visual changes in the 2D organ labels guide the training to predict the noise. Experimental results demonstrate that the prediction accuracy surpasses that of conventional approaches, producing image overlays that visualize tumors and vascular structures for intraoperative guidance.Clinical Relevance: The proposed 2D/3D registration model generates image overlays that facilitate the visualization of tumors and vascular structures in laparoscopic surgery, thereby enabling highly precise surgical guidance.
This study presents an automatic segmentation algorithm for accurately delineating lungs with pathological attenuations in 3D computed tomography (CT) images. Classical image processing methods are applied, including intrinsic image decomposition (IID) filtering and wavelet transform. Two contour refinement strategies—convex hull and corner detection—are evaluated, with the convex hull approach demonstrating superior performance, achieving a Dice similarity coefficient (DSC) of 98% against expert manual delineations. When compared to a state-of-the-art deep learning (DL) model (TotalSegmentatorV2), the classical approach remains competitive, achieving lower Hausdorff distance (HD), indicating fewer extreme segmentation errors. These results suggest that while DL methods provide high segmentation accuracy, classical approaches that combine 2D and 3D processing still offer advantages in mitigating outlier errors and ensuring interpretability. Additionally, the robustness and interpretability of classical methods make them ideal for generating accurate, well-annotated training datasets for DL models, enhancing performance, especially in lung disease contexts. Future work could explore hybrid models that leverage the strengths of both classical and DL-based techniques for robust lung segmentation.
Medical image classification plays an increasingly vital role in identifying various diseases by classifying medical images, such as X-rays, MRIs and CT scans, into different categories based on their features. In recent years, deep learning techniques have attracted significant attention in medical image classification. However, it is usually infeasible to train an entire large deep learning model from scratch. To address this issue, one of the solutions is the transfer learning (TL) technique, where a pre-trained model is reused for a new task. In this paper, we present a comprehensive analysis of TL techniques for medical image classification using deep convolutional neural networks. We evaluate six pre-trained models (AlexNet, VGG16, ResNet18, ResNet34, ResNet50, and InceptionV3) on a custom chest X-ray dataset for disease detection. The experimental results demonstrate that InceptionV3 consistently outperforms other models across all the standard metrics. The ResNet family shows progressively better performance with increasing depth, whereas VGG16 and AlexNet perform reasonably well but with lower accuracy. In addition, we also conduct uncertainty analysis and runtime comparison to assess the robustness and computational efficiency of these models. Our findings reveal that TL is beneficial in most cases, especially with limited data, but the extent of improvement depends on several factors such as model architecture, dataset size, and domain similarity between source and target tasks. Moreover, we demonstrate that with a well-trained feature extractor, only a lightweight feedforward model is enough to provide efficient prediction. As such, this study contributes to the understanding of TL in medical image classification, and provides insights for selecting appropriate models based on specific requirements.
