Detection and severity classification of Diabetic Retinopathy from Fundus Images Utilizing Deep Transfer Learning

Diabetic retinopathy (DR) is a primary cause of vision impairment among adults. The disease predominantly affects nations with intermediate or low economic status. Advancements in medical research offer methods to avoid vision loss caused by diabetic retinopathy; nevertheless, early diagnosis of the condition is essential and labor-intensive. This paper presents a computer-automated approach that precisely detects and classifies diabetic retinopathy from fundus images, thereby minimizing the need for human intervention. We have developed a bespoke dataset derived from Aptos, DDR, IDRiD, and photos from several publicly available datasets, which have been segmented and classified by expert pathologists. The dataset is treated using a Gaussian filter and histogram normalization for image enhancement. We also employed the Synthetic Minority Oversampling Technique (SMOTE) to equilibrate the dataset and mitigate bias issues. We have integrated three deep transfer learning models: ResNet152V2, DenseNet121, and InceptionV3. Additionally, we have constructed a bespoke lightweight deep learning model that functions effectively with a minimal number of photos. We furthermore implemented the Ensemble procedure utilizing soft voting to optimize the outcomes from the models. Gradient-Weighted Class Activation Mapping (GRAD-CAM) is also utilized to monitor the model’s convergence. Among all the models, ResNet152V2 exhibited the highest performance, with an accuracy of 99.52%. The ensemble technique attains optimal findings with an accuracy of 99.98%, enabling a dependable diagnosis of DR.