A multi-class fundus disease classification framework based on self-supervised contrastive learning and dual-branch network
- 1School of Medicine and Health, Guangzhou Xinhua University, Guangzhou, China
- 2School of Information Engineering, Xinyang Agriculture and Forestry University, Xinyang, China
Abstract
Introduction: Fundus diseases including diabetic retinopathy, glaucoma, and age-related macular degeneration are primary causes of blindness worldwide. Manual interpretation of retinal images is time-consuming and heavily reliant on clinical expertise. Although deep learning has shown promise in automated fundus image analysis, its performance remains highly data-dependent, requiring large amounts of expert annotations. Moreover, single-architecture models may have limited capacity to capture the diverse pathological features present in fundus images.
Methods: This study proposes a unified framework integrating self-supervised contrastive pre-training with a hybrid dual-branch architecture. Pre-training on unlabeled fundus images enables the model to learn generalizable representations, thereby reducing reliance on manual annotations. The dual-branch design combines EfficientNet-B4 for fine-grained detail extraction and MambaOut for large receptive field modeling. Their outputs are adaptively fused via a learnable cross-fusion module. A dendritic classifier is adopted to further improve classification performance. Comprehensive evaluations were conducted on two public datasets, OIHC and FIVES.
Results: On the OIHC dataset, the proposed model achieved an F1-score of 96.32 ± 0.38% and a Kappa coefficient of 95.89 ± 0.43%. On the FIVES dataset, it attained 95.10 ± 0.54% and 93.49 ± 0.71%. The macro-average AUC exceeded 98% on both datasets. Ablation studies demonstrated the contribution of each component to the overall performance.
Conclusion: This study indicates that combining self-supervised pre-training with a hybrid dual-branch architecture is a feasible approach for multi-class fundus disease classification. Interpretability analyses suggest that the model's attention regions partially overlap with clinically relevant pathological features. The framework achieved competitive performance on both datasets, showing its potential utility in scenarios with limited annotated data. Future work will focus on multi-center validation and model compression for lightweight architecture design.
Related
- Deep Learning for Multi‐Disease OCT Classification in Clinical Retina Practice Using Attention Mechanisms and Contrastive Learning
- Robust and generalizable atrial fibrillation detection from ECG using time-frequency fusion and supervised contrastive learning
- KT-CBHNET: a spatio-temporal deep learning framework for interpretable ECG signal classification for cardiovascular disease prediction
- An Explainable Disease-Aware Intelligent Histopathological Representation Learning Framework for Generalizable Breast Cancer Classification
- An uncertainty-aware multimodal deep learning framework for Alzheimer’s disease classification using MRI and electronic health records
BibTeX
@article{cai2026multiclass,
title = {{A multi-class fundus disease classification framework based on self-supervised contrastive learning and dual-branch network}},
author = {Lili Cai and Keke Hou and Zhong Mo and Shining Ding},
journal = {Frontiers in Medicine},
year = {2026},
month = oct,
doi = {10.3389/fmed.2026.1932626},
url = {https://doi.org/10.3389/fmed.2026.1932626},
note = {CC BY 4.0},
}
RIS
TY - JOUR
TI - A multi-class fundus disease classification framework based on self-supervised contrastive learning and dual-branch network
AU - Lili Cai
AU - Keke Hou
AU - Zhong Mo
AU - Shining Ding
PY - 2026
DA - 2026/10/09
JO - Frontiers in Medicine
DO - 10.3389/fmed.2026.1932626
UR - https://doi.org/10.3389/fmed.2026.1932626
AB - Introduction: Fundus diseases including diabetic retinopathy, glaucoma, and age-related macular degeneration are primary causes of blindness worldwide. Manual interpretation of retinal images is time-consuming and heavily reliant on clinical expertise. Although deep learning has shown promise in automated fundus image analysis, its performance remains highly data-dependent, requiring large amounts of expert annotations. Moreover, single-architecture models may have limited capacity to capture the diverse pathological features present in fundus images. Methods: This study proposes a unified framework integrating self-supervised contrastive pre-training with a hybrid dual-branch architecture. Pre-training on unlabeled fundus images enables the model to learn generalizable representations, thereby reducing reliance on manual annotations. The dual-branch design combines EfficientNet-B4 for fine-grained detail extraction and MambaOut for large receptive field modeling. Their outputs are adaptively fused via a learnable cross-fusion module. A dendritic classifier is adopted to further improve classification performance. Comprehensive evaluations were conducted on two public datasets, OIHC and FIVES. Results: On the OIHC dataset, the proposed model achieved an F1-score of 96.32 ± 0.38% and a Kappa coefficient of 95.89 ± 0.43%. On the FIVES dataset, it attained 95.10 ± 0.54% and 93.49 ± 0.71%. The macro-average AUC exceeded 98% on both datasets. Ablation studies demonstrated the contribution of each component to the overall performance. Conclusion: This study indicates that combining self-supervised pre-training with a hybrid dual-branch architecture is a feasible approach for multi-class fundus disease classification. Interpretability analyses suggest that the model's attention regions partially overlap with clinically relevant pathological features. The framework achieved competitive performance on both datasets, showing its potential utility in scenarios with limited annotated data. Future work will focus on multi-center validation and model compression for lightweight architecture design.
SN - 2296-858X
ER -