Related Experiment Videos
Automatic Classification of Retinal OCT Images Based on Multi-Perspective Collaborative Self-Distillation and
IEEE Journal of Biomedical and Health Informatics
|July 31, 2026
Summary
A new network, MCSD-Net, improves optical coherence tomography (OCT) image classification for diagnosing fundus diseases. It tackles class imbalance and feature similarity, enhancing diagnostic accuracy for retinal conditions.
Area of Science:
- Ophthalmology
- Medical Imaging
- Artificial Intelligence
Background:
- Automatic classification of retinal optical coherence tomography (OCT) images aids fundus disease diagnosis.
- Challenges include class imbalance and high inter-class similarity in OCT datasets.
Purpose of the Study:
- To propose a novel Multi-Perspective Collaborative Self-Distillation Network (MCSD-Net) for enhanced OCT image classification.
- To address class imbalance and improve feature discriminability in OCT image analysis.
Main Methods:
- Developed MCSD-Net incorporating a multi-perspective collaborative self-distillation mechanism (structural and historical distillation).
- Implemented a category-aware contrastive learning (CACL) strategy for balanced batches and robust feature representation.
- Integrated a direction-aware attention module (DAAM) to enhance feature discriminability.
Main Results:
- MCSD-Net demonstrated superior performance on private MTM and public OCT datasets (OCT2017, OCTDL).
- The proposed methods effectively alleviated class imbalance and improved feature learning.
- Outperformed existing state-of-the-art self-distillation-based methods in OCT image classification.
Conclusions:
- MCSD-Net offers a promising approach for accurate OCT image classification.
- The combination of self-distillation and contrastive learning effectively handles challenges in OCT data.
- This network can significantly assist in the early diagnosis of retinal diseases.