Enhancing Diabetic Retinopathy Identification Through Novel Image Representation and Hybrid CNN-Transformer Segmentation

Diabetic retinopathy, an insidious complication arising from diabetes, emerges as a predominant catalyst for vision impairment. Early identification of this condition remains a pressing need to prevent the risk of blindness. Many existing approaches for diabetic retinopathy identification focus on the processing of retinal images via standard preprocessing techniques before feeding them to classification models, without exploring alternative representations. Additionally, recent segmentation techniques, which may particularly contribute to diabetic retinopathy detection or progression, face challenges related to high computational and time consumption. To address diabetic retinopathy diagnosis from performance improvement and computational time reduction perspectives, we have focused on two primary contributions. The first involves diabetic retinopathy identification utilizing a novel fundus retinal image transformation approach. This novel representation aims to extract meaningful retinal information while discarding insignificant borders, effectively reducing image size. The second contribution entails retinal part segmentation using a hybrid approach that leverages the strengths of both convolutional neural networks and Vision Transformers in terms of rapidity and accuracy, respectively. The results obtained by the diabetic retinopathy identification using the novel representation approach on both datasets, namely EyePACS [5] and APTOS [6], as well as those obtained through the proposed segmentation approach on the IDRiD Dataset [9], are promising.