Cross-Modality Anatomy Knowledge Distillation for Retinal OCT Layer Segmentation With Limited Labels

Le résumé fourni par la source

Automatic segmentation of retinal Optical Coher ence Tomography (OCT) images plays a vital role in the early diagnosis and treatment of ocular diseases. However, the scarcity of labeled OCT data poses a significant challenge for existing supervised learning methods. In contrast, Color Fundus Photography (CFP) provides clearer visualization of key anatomical regions, such as the optic disc and cup, making it easier to acquire high-quality annotations. To address the issue of insufficient OCT annotations, we propose a Cross-Modality Anatomy Knowledge Distillation (CMAKD) framework. This method leverages a large amount of annotated CFP data to learn cross-modality structural priors related to retinal layer organization and vessel-associated landmarks, which are then used to regularize the OCT model for retinal layer segmentation. In other words, CFP serves as a source of weak structural priors, providing spatial and contextual guidance for OCT feature learning rather than depth-resolved layer supervision or a direct supervisory signal. To achieve this goal, we design a Domain-invariant Contrastive Learning (DoCL) module to reduce the modality gap between CFP and OCT by aligning their feature distributions. Furthermore, we introduce a Layer-aware Self-ensembling Mean-Teacher (LA-MT) framework to fully exploit unlabeled OCT data and enhance the model’s ability to recognize layered boundary structures. Experimental results on three public datasets — Duke DME, MS, and GOALS — demonstrate that CMAKD significantly improves segmentation performance while reducing the reliance on OCT annotations. Under semi-supervised domain adaptation settings, CMAKD achieves relative performance gains of 17.8% and 8.2% on the Duke DME and MS datasets, respectively, and approaches the performance of fully supervised models on the GOALS dataset. Our code is available at: https://github.com/lixiangcog/CMAKD.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.