Abstract:  Objective To establish an artificial intelligence (AI)-based diabetic retinopathy (DR) lesion segmentation system, compare its performance with traditional manual annotation, and explore its application in clinical teaching. Design Educational research study. Participants Ten first-year professional master's students from Beijing Tongren Hospital and 300 colored DR fundus photographs. Methods An AI-based fundus lesion segmentation system was established to identify microaneurysms (MA), hemorrhages (HE), hard exudates (EX), and soft exudates (SE). All fundus photographs were annotated by senior retinal disease experts, serving as the gold standard. Both AI and manual annotations were conducted for comparison. Two groups of ophthalmology students were taught using either the manual or AI method, and their test accuracy was compared. Main Outcome Measures The accuracy of lesion labeling in DR fundus photos (Dice coefficient), the time required to label each fundus photo, and the accuracy of the students' identification of focal lesions. Results The average Dice coefficients of AI and manual labeling were 0.845 and 0.871, respectively (P=0.525). The average time required by AI for each fundus photo was 1.2 seconds, which was significantly shorter than 1451 seconds for manual labeling(P<0.001). After teaching, with the two methods respectively, the average accuracy of students, assessment was 83.5% and 82.5% respectively (P=0.790). Conclusion The AI system exhibits robust performance in segmenting fundus lesions of DR, achieving comparable accuracy to manual annotation with significantly higher efficiency, which proves it can be applied in clinical education. (Ophthalmol CHN, 2025, 34: 232-235)

Key words: artificial intelligence, retina, lesion segmentation, ophthalmology, teaching