Abstract
BACKGROUND: Diabetic retinopathy (DR) is the leading cause of blindness among working-age adults. Early detection is crucial to reducing DR-related vision loss risk but is fraught with challenges. Manual detection is labor-intensive and often misses tiny DR lesions, necessitating automated detection. OBJECTIVE: We aimed to develop and validate an annotation-free deep learning strategy for the automatic detection of exudates and bleeding spots on color fundus photography (CFP) images and ultrawide field (UWF) retinal images. MATERIALS AND METHODS: Three cohorts were created: two CFP cohorts (Kaggle-CFP and E-Ophtha) and one UWF cohort. Kaggle-CFP was used for algorithm development, while E-Ophtha, with manually annotated DR-related lesions, served as the independent test set. For additional independent testing, 50 DR-positive cases from both the Kaggle-CFP and UWF cohorts were manually outlined for bleeding and exudate spots. The remaining cases were used for algorithm training. A multiscale contrast-based shape descriptor transformed DR-verified retinal images into contrast fields. High-contrast regions were identified, and local image patches from abnormal and normal areas were extracted to train a U-Net model. Model performance was evaluated using sensitivity and false positive rates based on manual annotations in the independent test sets. RESULTS: Our trained model on the independent CFP cohort achieved high sensitivities for detecting and segmenting DR lesions: microaneurysms (91.5%, 9.04 false positives per image), hemorrhages (92.6%, 2.26 false positives per image), hard exudates (92.3%, 7.72 false positives per image), and soft exudates (90.7%, 0.18 false positives per image). For UWF images, the model's performance varied by lesion size. Bleeding detection sensitivity increased with lesion size, from 41.9% (6.48 false positives per image) for the smallest spots to 93.4% (5.80 false positives per image) for the largest. Exudate detection showed high sensitivity across all sizes, ranging from 86.9% (24.94 false positives per image) to 96.2% (6.40 false positives per image), though false positive rates were higher for smaller lesions. CONCLUSIONS: Our experiments demonstrate the feasibility of training a deep learning neural network for detecting and segmenting DR-related lesions without relying on their manual annotations.