Diabetic Retinopathy is one of the common complications of diabetic patients, which can lead to vision loss or even blindness. Early detection and timely treatment are crucial for preventing vision loss. However, in real-world applications, DR classification tasks often face the challenge of imbalanced data distributions, where one class significantly outnumbers the others. Therefore, this project proposed a novel ensemble learning model which combines ResNet50 and a dual attention method for Diabetic Retinopathy classification, trained on the pre-processed APTOS 2019 Blindness Detection dataset downloaded from Kaggle. The model outperformed the pre-training model with an Average Precision (AP) of 1 in the No DR class and an Area Under the Curve (AUC) of 0.97 in the Advanced class, 0.93 in the Mild class, 1 in the No DR class, and 0.96 in the Proliferative DR class. Additionally, this project utilizes Local Interpretable Model-Agnostic Explanations (LIME) as the interpretability tool. In conclusion, the integrated model combining ResNet50 with the dual attention strategy performs admirably in the diabetic retinopathy classification task. It holds strong potential for auxiliary medical diagnosis, demonstrating both robust discriminative ability and excellent interpretability under imbalanced data scenarios.
This project, spearheaded by Zhao Yikai (Richard), addresses a critical need in medical imaging—early and accurate classification of Diabetic Retinopathy (DR). Given the challenges of imbalanced datasets and subtle retinal variations, the team developed a novel Residual-Based Dual Attention Ensemble model integrating ResNet50 and spatial-channel attention mechanisms.
The model demonstrates strong classification performance across multiple DR categories and offers explainable insights into prediction logic. Its accuracy and interpretability make it a valuable tool for aiding clinical decisions in real-world settings. The visualizations provided by LIME confirmed that the model focused on medically relevant retinal regions.