Delayed treatment of infantile retinal disease can reduce its effectiveness and may cause severe and irreversible damage. Automated diagnosis of infant retinal diseases faces challenges including subtle early lesions, diverse clinical phenotypes, imaging variations, and imbalanced data. To address these, which cannot be well addressed by existing general foundation models, we propose structure-aware bidirectional proxy interaction network (SABPI-Net) in a universal learning framework. SABPI-Net incorporates a high-frequency mapping branch, and employs a proposed proxy interaction attention module to enable effective interaction between its trunk feature encoding branch and the high-frequency mapping branch, thereby facilitating enhanced perception of retinal detail structures. Domain-agnostic embedding space self-matching, guided by a memory-bank low-frequency component replacement strategy, promotes domain-invariant learning and consistent model performance under diverse image styles. Finally, the tail-aware feature fusion strategy for fine-tuning further enhances the model's diagnostic sensitivity to tailed diseases. In this study, three classification tasks related to infant retinal diseases are implemented on the largest clinical infant retina dataset to date, covering 19 infant retinal diseases or normal conditions. SABPI-Net achieves superior performance compared to 13 SOTA methods, with 95.32% accuracy on mainstream clinical tasks, 73.58% on ROP five-stage classification, and 84.25% on multi-disease classification, representing improvements of 1.57%, 1.88%, and 4.71% respectively over the best competing methods. Extensive experiments demonstrate the effectiveness and superiority of SABPI-Net in diagnosing infant retinal diseases.

Fundus image classification plays a crucial role in diagnosing ophthalmic diseases but remains challenging due to the scarcity of annotated data and the subtlety of lesion features, which often resemble surrounding tissues. Vision-language models (VLMs), known for their impressive few-shot learning performance on natural images, offer a promising foundation for medical image analysis. However, directly applying these models to medical tasks is suboptimal, as they lack domain-specific knowledge and struggle to capture fine-grained pathology cues. To address these challenges, we propose a novel framework for few-shot fundus image classification that integrates medical knowledge-driven prompt learning into CLIP. Specifically, our method utilizes a domain-specific prompt bank constructed from clinical terminology to enrich the model's understanding of medical context. Additionally, we introduce a cross-modal alignment loss to improve consistency between visual and textual features and employ a lightweight adapter for efficient task-specific fine-tuning. Extensive experiments across multiple datasets demonstrate that our approach significantly enhances performance, surpassing existing methods in various few-shot scenarios.

Retinopathy of prematurity (ROP) is a retinal vascular disease that primarily affects premature infants with low birth weight. It is a leading cause of childhood blindness worldwide, but it can often be effectively managed with appropriate and timely diagnosis and treatment. To address the impact of image style on model classification performance, this paper proposes a dual-scale Swin Transformer (DS-Swin-T) network for ROP. The network comprises three components: image synthesis (IS), feature alignment, and advanced adversarial learning. The IS module generates synthesis style images as an intermediate latent space between source and target styles, reducing style difference. The DS-Swin-T serves as the primary framework for image feature extraction. Detail and style encoders extract features in the shallow feature space, with detail and style losses aligning these features to ensure consistency across styles. To extract rich style-invariant features and ensure consistent classification within the same category, adversarial learning is applied in the advanced feature space. Finally, feature fusion units process dual-scale classification representations. Our method achieves an average accuracy of 97.91% on the source style dataset. When transferred to other target style datasets, our method effectively mitigates the performance degradation caused by style difference, reaching a maximum average accuracy of 93.66%. Extensive experiments demonstrate the effectiveness of our method.

Automatic and accurate classification of retinal optical coherence tomography (OCT) images is essential to assist physicians in diagnosing and grading pathological changes in pathologic myopia (PM). Clinically, due to the obvious differences in the position, shape, and size of the lesion structure in different scanning directions, ophthalmologists usually need to combine the lesion structure in the OCT images in the horizontal and vertical scanning directions to diagnose the type of pathological changes in PM. To address these challenges, we propose a novel feature interaction Transformer network (FIT-Net) to diagnose PM using OCT images, which consists of two dual-scale Transformer (DST) blocks and an interactive attention (IA) unit. Specifically, FIT-Net divides image features of different scales into a series of feature block sequences. In order to enrich the feature representation, we propose an IA unit to realize the interactive learning of class token in feature sequences of different scales. The interaction between feature sequences of different scales can effectively integrate different scale image features, and hence FIT-Net can focus on meaningful lesion regions to improve the PM classification performance. Finally, by fusing the dual-view image features in the horizontal and vertical scanning directions, we propose six dual-view feature fusion methods for PM diagnosis. The extensive experimental results based on the clinically obtained datasets and three publicly available datasets demonstrate the effectiveness and superiority of the proposed method. Our code is avaiable at: https://github.com/chenshaobin/FITNet.