ISGM-DR: Important Slice-Guided Model for Diabetic Retinopathy Diagnosis Using 3D Oct Images

Early detection of diabetic retinopathy (DR) is essential for preventing irreversible vision loss. This study introduces the Important Slice-Guided Model for DR Diagnosis (ISGM-DR), which utilizes 3D-Optical Coherence Tomography (OCT) images to effectively differentiate between normal retinas and those affected by DR. As the name suggests, we propose that within the 3D OCT images lies a critical slice, guiding ophthalmologists to swift diagnosis and action upon detecting abnormalities. Accordingly, our model is composed of two key components: the Adaptive Importance Slice-based Network (AISN) and the Channel-Aware Information Slice Block (CAISB). The AISN generates a probability map for each slice, quantifying the importance of each slice in contributing to the diagnosis of DR. Based on these probability maps, we identify the most informative slice from the 3D OCT images. Furthermore, to extract additional evidence regarding the semantic cues associated with DR, we apply the CAISB to obtain highly attended slices enriched with information pertinent to the disease. To propagate this information to the output of the backbone network, we perform element-wise multiplication on the feature maps of the backbone and the highly informative slice. This process guides the entire model toward accurate DR diagnosis. Evaluating a diverse cohort of 481 3D-OCT volumes from both healthy and diseased subjects, our ISGM-DR model achieves an impressive overall accuracy of 93.78%, with a recall of 85.71% and a precision of 92.31%. This performance significantly surpasses existing state-of-the-art methods for DR classification, including Vision Transformer (ViT), Max ViT, and other baseline CNN models.