Title: HD-DinoMoE: A Class-Aware Hierarchical Dual Mixture-of-Experts Network for Scleral Anomaly Segmentation in Complex Acquisition Scenarios

Abstract:

While Traditional Chinese Medicine (TCM) ocular inspection offers empirical indicators for evaluating scleral surface anomalies, its application in clinical settings is often hindered by subjectivity and a lack of quantifiable metrics. To facilitate intelligent and measurable ocular examination, this research introduces the TCM-inspired Artificial Intelligence Ocular Auxiliary Diagnosis System (TAO), with a specific emphasis on pixel-level segmentation of scleral surface irregularities. Addressing challenges inherent in images captured by both clinical and user sources—such as multi-source distributional discrepancies, varied anomaly morphologies, and scleral specular reflection (SSR)—we propose HD-DinoMoE. This is a class-aware hierarchical dual mixture-of-experts network designed to segment vessels, yellow spots, black spots, and blood spots.

HD-DinoMoE integrates class-aware dual-stream DINOv3 feature fusion with class-specific multi-expert decoding. To ensure robust performance, the system employs a three-stage backbone-frozen routing strategy to stabilize the adaptation of dual backbones. Furthermore, the Progressive Confidence Penalty (PCP) Loss is utilized to mitigate high-confidence false positives and segmentation leakage within SSR regions, while Class-Aware Adaptive Sample Weighting (CA-ASW) balances training contributions at both the sample and class levels.

The study also introduces the Multi-label Scleral Anomaly Segmentation Dataset (ML-SASD), a novel benchmark featuring pixel-wise annotations for three anomaly categories across Clinical, Wild, and Mix settings. Evaluations on the ML-SASD-Mix subset demonstrate that HD-DinoMoE achieves a mean Dice score of 72.11% and a mean Intersection-over-Union (mIoU) of 58.44%, while effectively maintaining boundary localization and controlling false positives in specular regions. Additionally, the model exhibits competitive generalization capabilities on the Vessels subset of the public SBVPI dataset. These findings suggest that HD-DinoMoE offers a viable segmentation solution for TAO in complex acquisition environments. Access to the code and data is provided at https://github.com/FX-CMX/HD-DinoMoE.

Source: arXiv Generated at: 2026-06-04 00:00:00 UTC