Xin Long,1– 5,* Fan Gan,1– 4,* Huimin Fan,1– 4,* WeiGuo Qin,6 Xiaonan Li,1– 4 Rui Ma,1 Leran Wang,1 Rui Hu,1 Yilin Xie,1 Lei Chen,7 Jian Cao,1– 4 Yinan Shao,1– 4 Kangcheng Liu,1– 4 Zhipeng You1– 4 

1Department of Fundus Diseases, The Affiliated Eye Hospital, Jiangxi Medical College, Nanchang University, Nanchang, Jiangxi, 330000, People’s Republic of China; 2Jiangxi Province Division of National Clinical Research Center for Ocular Diseases, Nanchang, Jiangxi, 330000, People’s Republic of China; 3Jiangxi Clinical Research Center for Ophthalmic Disease, Nanchang, Jiangxi, 330000, People’s Republic of China; 4Jiangxi Research Institute of Ophthalmology and Visual Science, Nanchang, Jiangxi, 330000, People’s Republic of China; 5School of Mathematics and Computer Sciences, Nanchang University, Nanchang, Jiangxi, 330000, People’s Republic of China; 6Department of Cardiothoracic Surgery, The 908th Hospital of Chinese People’s Liberation Army Joint Logistic Support Force, Nanchang, Jiangxi, 330001, People’s Republic of China; 7The Second Clinical School of Medicine, Nanchang University, Nanchang, Jiangxi, 330031, People’s Republic of China

*These authors contributed equally to this work

Correspondence: Zhipeng You, The Affiliated Eye Hospital, Jiangxi Medical College, Nanchang University, 463 Bayi Road, Nanchang, Jiangxi, 330006, People’s Republic of China, Tel +86 13970032654, Email yzp74@sina.com

Purpose: This study aims to develop and validate a deep learning-based automated diagnostic system that utilizes fluorescein angiography (FFA) images for the rapid and accurate diagnosis of diabetic retinopathy (DR) and its complications.
Methods: We collected 19,031 FFA images from 2753 patients between June 2017 and March 2024 to construct and evaluate our analytical framework. The images were preprocessed and annotated for training and validating the deep learning model. The study employed a two-stage deep learning system: the first stage used EfficientNetB0 for a five-class classification task to differentiate between normal retinal conditions, various stages of DR, and post-laser treatment status; the second stage focused on images classified as abnormal in the first stage, further detecting the presence of diabetic macular edema (DME). Model performance was evaluated using multiple classification metrics, including accuracy, AUC, precision, recall, F1-score, and Cohen’s kappa coefficient.
Results: In the first stage, the model achieved an accuracy of 0.7036 and an AUC of 0.9062 on the test set, demonstrating high accuracy and discriminative ability. In the second stage, the model achieved an accuracy of 0.7258 and an AUC of 0.7530, performing well. Additionally, through Grad-CAM (gradient-weighted class activation mapping), we visualized the most influential image regions in the model’s decision-making process, enhancing the model’s interpretability.
Conclusion: This study successfully developed an end-to-end DR diagnostic system based on the EfficientNetB0 model. The system not only automates the grading of FFA images but also detects DME, significantly reducing the time required for image interpretation by clinicians and providing an effective tool to improve the efficiency and accuracy of DR diagnosis.
Plain Language Summary: This study developed an automated deep learning-based diagnostic system for diabetic retinopathy (DR) and diabetic macular edema (DME) using fluorescein angiography (FFA) images, enabling rapid and accurate diagnosis. We collected 19,031 FFA images from 2753 patients and trained a two-stage deep learning model. In the first stage, the system used the EfficientNetB0 model for a five-class classification task, distinguishing between normal retinal conditions, various DR stages, and post-laser treatment status. The model achieved an accuracy of 70.36% and an AUC of 0.9062 on the test set. In the second stage, the model focused on detecting DME in images classified as abnormal in the first stage, with an accuracy of 72.58% and an AUC of 0.7530. Additionally, Grad-CAM visualization highlighted key image regions influencing the model’s decision, enhancing its interpretability. This system not only automates FFA image grading but also detects DME, significantly reducing the time needed for clinicians to interpret images, and improving the efficiency and accuracy of DR diagnosis. It holds great potential for widespread clinical application.

Keywords: diabetic retinopathy, DR, diabetic macular edema, DME, deep learning, EfficientNetB0, fluorescein angiography, FFA