RetinoNet-VT: A Hybrid Deep Learning Architecture for Diabetic Retinopathy Lesion Segmentation

Diabetic retinopathy (DR) is a leading cause of blindness worldwide. It is diagnosed by identifying key retinal lesions, including intraretinal microvascular abnormalities (IR-MAs), non-perfusion areas (NPAs), and neovascularization (NV). Ultra-wide optical coherence tomography angiography (UW-OCTA), with its high-resolution and extensive field of view, has emerged as a powerful imaging modality for detecting lesions associated with DR. However, accurate segmentation of these lesions remains a challenge due to their complex structures, variations in size and shape, and low contrast in UW-OCTA images. While deep learning-based segmentation approaches show promise in automating DR lesion detection, their accuracy remains limited due to inadequate feature representation, restricted receptive fields, and difficulty in capturing long-range dependencies. To address this, we propose RetinoNet-VT, a novel deep-learning architecture that builds upon the UNet framework for DR lesion segmentation. RetinoNet-VT integrates three key components: a pre-trained VGG16 encoder for hierarchical feature extraction, a multi-head self-attention transformer in the bottleneck to capture long-range dependencies, and deep supervision at multiple decoder levels to improve gradient propagation and multi-scale feature learning. We evaluate our model on the DRAC 2022 dataset, achieving F1 scores of 0.6596 for IRMA, 0.8359 for NPA, and 0.7985 for NV, outperforming existing methods. These results highlight the superior accuracy of RetinoNet-VT, underscoring its potential for integration into automated clinical workflows for early DR detection and effective disease monitoring.