Cataract remains one of the leading causes of blindness and visual impairment worldwide, representing a significant public health concern. Anterior segment optical coherence tomography (AS-OCT) provides high-resolution visualization of ocular structures and has become a key imaging modality for nuclear cataract (NC) grading. However, existing convolutional neural network (CNN)-based methods often struggle to differentiate subtle variations between adjacent severity levels due to limited capacity for capturing long-range dependencies, thereby affecting classification accuracy. To address this challenge, we propose an automatic nuclear cataract grading network based on the Mamba architecture. This framework combines the local feature extraction capabilities of traditional CNNs with the long-range dependency modeling power of Mamba modules. Furthermore, we introduce a Hybrid Wavelet Feature Refinement Module (HWFRM), which employs wavelet transforms to extract multi-frequency representations. Integrated with a detail-guided enhancement mechanism, the module adaptively strengthens discriminative features. Channel and spatial attention mechanisms are applied to each wavelet sub-band, enabling the network to selectively emphasize important frequency components and remain sensitive to both structural and fine-detail cues. Finally, an ordinal regression loss is incorporated to explicitly model the progressive nature of cataract severity, improving the network’s ability to reduce misclassifications between adjacent categories. Extensive experiments on both a local AS-OCT dataset and a public benchmark demonstrate that our approach achieves state-of-the-art performance.

Accurate segmentation of key anatomical structures in anterior segment OCT (AS-OCT) images is critical for diagnosing serious ophthalmic conditions such as keratitis and cataract. However, due to the scarcity of labeled data in this domain, most existing methods struggle to precisely segment both the lens and the anterior chamber angle simultaneously. To address these limitations, we propose a semi-supervised segmentation framework based on collaborative training between U-Net and Mamba-UNet. A Scale Fusion Module (SFM) is introduced to integrate the outputs of both models, generating multi-scale predictions and fused pseudo-labels. A multi-scale supervision strategy is then employed to guide learning at different levels. Additionally, we design a novel anatomical structure consistency loss that leverages anatomical properties from the fused pseudo-labels to preserve anatomical correctness. Experimental results on two AS-OCT datasets demonstrate the effectiveness and superiority of our proposed approach.

Accurate structural segmentation and landmark detection in anterior segment optical coherence tomography (AS-OCT) images are crucial for extracting clinical parameters that guide the diagnosis and treatment of diseases such as glaucoma. However, current mainstream algorithmic paradigms suffer from an inherent limitation: their performance improvements heavily rely on large amounts of high-quality annotations. To overcome this bottleneck, we propose a novel semi-supervised multi-task learning framework. Our framework first incorporates the powerful Segment Anything Model (SAM) image encoder to enhance the model’s general feature extraction capability. To address SAM’s adaptability issues in the medical imaging domain, we design an adaptive feature fusion adapter (AFFA) for targeted fine-tuning, thereby improving its performance on AS-OCT images. Simultaneously, our proposed synergistic feature exchange module (SFEM) enables mutual promotion between the segmentation and detection tasks. Experimental results on a local dataset demonstrate that our proposed method achieves superior performance.

Optical coherence tomography (OCT) is an important ophthalmic imaging technique, which can generate high-resolution anatomical images and plays an important role in the detection of retinal biomarkers. However, the appearance of retinal biomarkers is complex, and some of these biomarkers differ greatly among different categories, while many features are similar. In addition, the boundaries of retinal biomarkers are often indistinguishable from the background. In this study, we propose a self-supervised contrastive boundary consistency network (SCB-Net) to detect retinal biomarkers in OCT images. A self-supervised contrastive classification module is proposed to improve the classification ability of the network between different categories of retinal biomarkers. Furthermore, in order to make the boundary of the retinal biomarkers located by the network closer to the ground truth, the boundary consistency is added on the basis of the original regressor to jointly constrain the boundary localization. The experimental results on a local dataset show that our proposed SCB-Net method achieves good detection performance compared with other detection methods.

Infrared imaging is currently the most effective clinical method to evaluate the morphology of the meibomian glands (MGs) in patients. As an important indicator for monitoring the development of MG dysfunction, it is necessary to accurately measure gland-drop and gland morphology. Although there are existing methods for automatic segmentation of MGs using deep learning frameworks, they require fully annotated ground-truth labels for training, which is time-consuming and laborious. In this article, we proposed a new scribble-supervised deep learning framework for segmenting the MGs, which only requires easily attainable scribble annotations for training. To cope with the shortage of supervision and regularize the network, a transformation consistent strategy is incorporated, which requires the prediction to follow the same transformation if a transform is performed on an input image of the network. The proposed segmentation method consists of two stages. In the first stage, a U-Net network is used to obtain the meibomian region segmentation map. In the second stage, we concentrate on segmenting glands in the meibomian region. We utilize the gradient prior information of the original image at the decoder part of the segmentation network, which can coarsely locate the target contour. We automatically generate reliable labels using the exponential moving average of the predictions during training and filter out the unreliable pseudo-label by uncertainty threshold. Experimental results on a local MG dataset and two other public medical image datasets demonstrate the effectiveness of the proposed segmentation framework.

When optical coherence tomography (OCT) is used for retinal disease diagnosis, it is critical to detect and classify the biomarkers from the OCT B-scans of patients. In this paper, we propose a novel weakly supervised approach that utilizes healthy data and image-level labels for biomarker detection and classification. The proposed approach is based on a hybrid network which integrates adversarial generative network and guided attention into one framework. The framework includes an anomaly detection network and a classification network. The anomaly detection network reconstructs an input image with biomarkers to a reconstructed image and the reconstructed image is compared with the input image to locate the biomarkers. Inspired by the guided attention inference network, we utilize the discriminator trained in the anomaly detection network as a classifier twice to reduce model parameters and obtain a complete attention map with class information to get biomarker classes. Experimental results with a large dataset demonstrate the effectiveness of the proposed detection and classification framework.

Near-infrared imaging is currently the most effective clinical method for evaluating the morphology of the meibomian glands in patients. Meibomian gland dysfunction (MGD) is a chronic and diffuse disease of the meibomian glands, which is an important cause of eye diseases such as dry-eye and blepharitis. Therefore, it is important to monitor the gland-drop and gland morphology for MGD patients. In this paper, we proposed a new scribble-supervised deep learning method for segmenting the meibomian glands. The proposed segmentation network consists of two stages. The first stage uses the U-Net network to obtain the meibomian region segmentation map. The second stage focuses on the meibomian region, combining spatial attention, gradient map and label filtering to generate the meibomian gland segmentation results. Experimental results on a local meibomian gland dataset demonstrate the effectiveness of the proposed segmentation framework.

Optical coherence tomography (OCT) imaging can obtain high-resolution cross-sectional scans of the retina, which can be used in clinical diagnosis. Changes in the thickness of layers indicate the onset of retinal diseases, motivating an accurate measurement of the thickness of retinal layers. Thus, an automatic and robust layer segmentation method is necessary. In this paper, we propose a deep learning-based multi-task framework to obtain the topologically consistent layer segmentation in OCT B-scans. By integrating the distance maps of retinal layer surfaces, the segmentation task is regarded as a multi-task problem of regression and classification. Besides, considering the multi-task learning problem, we propose a task-specific attention module to learn the task-tailored features. Experiment results on a public OCT dataset with multiple sclerosis (MS) demonstrate the effectiveness of the proposed method.