Related Experiment Video
Updated: Jul 30, 2026
07:41
Behavioral Assessment of Visual Function via Optomotor Response and Cognitive Function via Y-Maze in Diabetic Rats
Published on: October 23, 2020
6.6K
Weakly supervised object detection network for diabetic retinopathy
Xulin Zong1, Bingxue Liang2, Yuhua Qin2
1Faculty of Information Science and Engineering, Ocean University of China, Qingdao, China.
Medical Physics
|December 31, 2025
Summary
Diabetic retinopathy detection (DRD-Net) improves early diagnosis using weakly supervised learning and enhanced feature extraction. This method accurately identifies lesions, potentially boosting clinical screening efficiency for diabetic patients.
Area of Science:
- Medical Imaging
- Artificial Intelligence
- Ophthalmology
Background:
- Increasing prevalence of diabetic patients necessitates rapid and accurate early diabetic retinopathy diagnosis.
- Diabetic retinal lesion identification is challenging due to the need for specialist judgment across multiple image regions, making manual labeling costly and time-consuming.
Purpose of the Study:
- To develop the Diabetic Retinopathy Detection Network (DRD-Net), an improved weakly supervised object detection model.
- DRD-Net utilizes an adversarial complementary erasure learning (ACoL) framework to enhance small lesion localization using only image-level labels, reducing diagnostic costs.
Main Methods:
- DRD-Net incorporates an enhanced EfficientNet-B0 with parallel downsampling and an efficient channel attention (ECA) module for feature extraction.
- A multi-scale parallel attention module (MPA) combined with ACoL improves classification and localization of small lesions across multi-scale features.
- Utilized 35,828 re-annotated lesion patches (224x224 pixels) from three datasets, partitioned into training (70%), validation (20%), and test (10%) sets.
Main Results:
- DRD-Net outperformed state-of-the-art methods, achieving 82.41% Top-1 Classification Accuracy, 76.94% Top-1 Localization Accuracy, and 86.05% Ground-Truth Known Localization Accuracy.
- Demonstrated significant performance gains over top baselines: 1.64% in Top-1 Cls (p=0.0318), 3.16% in Top-1 Loc (p=0.0090), and 0.96% in GT-Known Loc (p=0.0481).
Conclusions:
- DRD-Net demonstrates feasibility for accurate and comprehensive diabetic retinopathy lesion identification.
- The model shows potential to enhance clinical screening efficiency and advance diabetic retinopathy detection technologies.
