Conv-ViT: An improved discrete convolution-based vision transformer for diabetic retinopathy detection

Abstract

Timely detection of Diabetic Retinopathy (DR), a major cause of irreversible blindness, is important to avert vision impairment. Present computer-aided diagnostic methods often suffer from poor segmentation, image noise, and a lack of generalization across datasets. This study proposes Conv-ViT, a hybrid model that integrates convolutional networks to overcome the disadvantages of the aforementioned models. Probability-based particle swarm optimization (PBPSO) was applied to achieve accurate segmentation, median filtering was applied to remove noise, and local binary pattern (LBP) was used to extract texture features. An innovative Electric Fish Optimization Arithmetic Algorithm (EFAOA), which enhances the compromise between exploration and exploitation during hyperparameter fine-tuning, was introduced to bolster model efficiency. Evaluation on the MESSIDOR dataset showed remarkable results, with an accuracy of 99.58 %, 98.86 %, 98.87 %, and an F1-score of 98.85 %, respectively. These results indicate the generalizability of the model beyond that of current state-of-the-art algorithms. The Conv-ViT framework represents a robust and scalable solution for the early detection of DR and holds great promise for use in automated cloud-based diagnostic systems.