Automated detection of proliferative & non-proliferative diabetic retinopathy using near-infrared and OCT imaging.

PURPOSE: Diabetic retinopathy, a major cause of blindness in working-age individuals, advances without adequate recognition and treatment from non-proliferative to proliferative diabetic retinopathy. Accurate classification of these stages is vital for timely intervention, though existing technologies still face challenges in part due to the variability of diabetic retinopathy features. This study investigates a multimodal deep learning model to automatically classify proliferative and non-proliferative diabetic retinopathy. METHODS: The model integrates optical coherence tomography and near-infrared imaging using a dataset of 1024 paired images from 226 patients. An advanced machine learning model, a so-called vision transformer with Masked Autoencoders, is used for multimodal learning. This method processes retinal images in segmented sequences to distinguish between structural details from coherence tomography and vascular patterns from infrared imaging. The model's effectiveness is evaluated through refined training and diagnostic sensitivity assessments, using the ROC-AUC metric to measure its accuracy in differentiating between non-proliferative and proliferative stages of diabetic retinopathy. RESULTS: Results indicate that the multimodal model significantly outperformed single-modality models, achieving a ROC-AUC score of 0.92, compared to 0.89 for optic coherence tomography alone and 0.87 for near-infrared images alone. CONCLUSION: The model's enhanced accuracy is attributed to the complementary nature of both modalities, with tomography providing structural details and near-infrared images highlighting vascular features.