Med Phys. 2026 Jan;53(1):
e70264
BACKGROUND: With the increasing number of diabetic patients, the rapid and accurate diagnosis of early diabetic retinopathy becomes crucial. However, diabetic retinal lesions are challenging to label since the identification of disease depends on accessing multiple lesion regions in images, which requires specialists to make detailed judgments and label them, resulting in an extremely costly and time-consuming.
PURPOSE: To reduce costs, we propose diabetic retinopathy detection network (DRD-Net), an improved weakly supervised object detection model based on adversarial complementary erasure learning (ACoL) framework, designed for diabetic retinopathy detection. DRD-Net enhances small lesion localization while relying only on image-level labels.
METHODS: DRD includes an improved EfficientNet-B0 network, which leverages compressed network structure with parallel downsampling and the efficient channel attention (ECA) module for feature extraction from fundus images. A multi-scale parallel attention module (MPA) is designed and combines with adversarial complementary erasure learning to enhance classification and localization of small lesion accross multi-scale features. For data processing, we cropped and re-annotated three datasets into 35,828 lesion patches (224 × 224 pixels) to solve the problem of information loss in high-resolution fundus images. The dataset is partitioned into training (25,079, 70%), validation (7166, 20%), and test (3583, 10%) sets. Benchmark models include CNN-based methods (CAM, ACoL, SLT-Net, etc.) and Transformer-based approaches (TRT, SAT, etc.). Performance is evaluated using Top-1/Top-5 classification accuracy (Top-1/Top-5 Cls), Top-1/Top-5 localization accuracy (Top-1/Top-5 Loc), and ground-truth known localization accuracy (GT-Known Loc). Statistical analyses employs paired t-tests with Holm-Bonferroni correction for multiple comparisons, Cohen's d for effect size, and a significance level of α = 0.05.
RESULTS: Experiments verify that the performance of DRD-Net is better than the state-of-the-art methods, achieving 82.41%, 76.94%, and 86.05% in Top-1 Cls, Top-1 Loc and GT-Known Loc, respectively. Compared to top-performancing baselines, gains are 1.64% (Cohen's d = 1.016, p = 0.0318), 3.16% (d = 1.377, p = 0.0090), and 0.96% (d = 0.85, p = 0.0481), all significant at α = 0.05.
CONCLUSIONS: Experiments confirm that DRD-Net has good feasibility to accurately and comprehensively identify DR lesions. This suggests that it could potentially enhance clinical screening efficiency and promote further development in diabetic retinopathy detection.
Keywords: deep learning; diabetic retinopathy detection; weakly supervised learning