Evaluation of classification performance for six types of fundus diseases in OCT images based on multi-source training strategy

TL;DR

This study improves OCT image classification for six fundus diseases by combining local and public datasets, enhancing model accuracy and reducing misdiagnosis rates.

Contribution

A novel multi-source training strategy is proposed to address class imbalance and limited data in OCT classification.

Findings

Combining local and public datasets (S2) significantly improved model performance compared to using public data alone (S1).

ViT-Base achieved 93.61% accuracy with reduced misdiagnosis rates for specific diseases like RAO and AMD.

Abstract

Currently, publicly available Optical Coherence Tomography (OCT) datasets are commonly plagued by limited coverage of disease categories, scarce samples and severe class imbalance, which leads to insufficient generalization ability of deep learning models in real-world clinical settings. This study aims to construct a high-quality OCT dataset encompassing six key types of fundus lesions and normal controls, and to systematically evaluate the improvement effect of training strategies for multi-source data fusion on the performance of multi-class classification.

We integrated local clinical data from Shanxi Eye Hospital with the latest public dataset OCTDL to establish a combined dataset. This dataset consists of 6,165 images, covering seven categories: age-related macular degeneration (AMD), diabetic macular edema (DME), retinal artery occlusion (RAO), retinal vein occlusion (RVO), epiretinal membrane (ERM), vitreomacular interface disease (VID), and normal controls (NO). On this basis, six representative deep learning architectures were selected, and two training paradigms were compared under unified experimental settings: (1) Training exclusively on open-source OCTDL data (S1); (2) Joint training using both local data and OCTDL data (S2). All models were evaluated on the identical OCTDL test set. A comprehensive analysis was conducted using multi-dimensional metrics including accuracy, weighted F1-score, class-specific recall, and area under the curve (AUC), with a particular focus on the misdiagnosis rate.

The S1 strategy exhibited significantly limited model recognition capability due to the extremely small sample sizes of certain categories. In contrast, the S2 strategy markedly improved the overall performance of the models. Confusion matrix analysis demonstrated that ViT-Base achieved the optimal performance under the S2 strategy: the accuracy reached 93.61%, the misdiagnosis rate of RAO was reduced to 0%, the misdiagnosis rate of AMD was controlled at 1.34%, and the misdiagnosis rate of RVO decreased from 14.89 to 8.51%.

Multi-source data fusion serves as an effective approach to enhance the robustness of OCT multi-category classification models, and it can notably strengthen the recognition capability for certain diseases in particular. This study not only verifies the universal benefits of this strategy but also reveals the critical impact of model selection on the transfer learning effect.

Figures