Automatic Classification of Retinal OCT Images Based on Multi-Perspective Collaborative Self-Distillation and...

The automatic classification of retinal optical coherence tomography (OCT) images holds significant value for assisting in the diagnosis of related fundus diseases. However, it often faces challenges such as class imbalance and high inter-class similarity. In this paper, we propose a multi-perspective collaborative self-distillation network (MCSD-Net) for OCT image classification, which enables the model to learn through multi-perspective collaborative self-distillation mechanism and category aware contrastive learning strategy. Specifically, a multi perspective collaborative self-distillation mechanism is designed, which comprises structural distillation based on multi-stage feature fusion and historical distillation based on linear growth strategy, enhancing knowledge diversity from both structural and historical perspectives. A category-aware contrastive learning (CACL) strategy is designed, which constructs class-balanced batches to alleviate class imbalance and employs contrastive learning to strengthen the representation of intra-class and inter class features. Additionally, a direction-aware attention module (DAAM) is incorporated to extract features from horizontal and vertical directions, thereby improving feature discriminability. The proposed MCSD-Net is evaluated on a private myopic tractional maculopathy (MTM) retinal OCT dataset and two public OCT datasets (OCT2017 and OCTDL). Experimental results demonstrate that the proposed MCSD Net outperforms other state-of-the-art self-distillation based methods.