Style-Prior Enhanced-and-Selective Mamba with Class-Oriented Contrastive Learning for Posterior Subcapsular Cataract Grading.

Abstract

Posterior subcapsular cataract (PSC) is a leading cause of blindness and vision impairment globally. Precise PSC grading is critical for immediate treatment and opacity progression observation. Recent studies have suggested the great potential of anterior segment optical coherence tomography (AS-OCT) images in objectively investigating the opacities of PSC compared to conventional slit-lamp examinations with clinical features such as average and maximum. However, the irregular morphology and diffuse distribution modeling of its opacities in AS-OCT images present two significant challenges for precise PSC grading, which have not been addressed previously. To address these challenges, we propose a novel Style-Prior Enhanced-and-Selective State Space Model (SPES-SSM), which effectively captures the irregular morphology and diffuse opacity distribution representations by leveraging the potential of SSM and multi-scale convolution concurrently; and then employs the well-designed style-prior enhanced-and-selective representation block to estimate the relative significances of these two representation types via useful clinical feature (e.g., average and maximum) prior exploration of PSC. By stacking SPES-SSM modules, we construct SPES-Mamba for automated PSC grading based on AS-OCT images. To enhance our SPES-Mamba's reliability and further improve its PSC grading results, we design a class-oriented contrastive learning (COCL), which explores the potential of the ranking matrix of label distance from the aspect of class-averaging. Extensive experiments demonstrate that SPES-Mamba and COCL significantly outperform state-of-the-art methods in terms of PSC grading results and reliability.