Transfer Learning for Five-Level Diabetic Retinopathy Severity Classification Using InceptionV3 and ResNet50

Tanuja Gannavarapu, Sai Prakash Gopalam, Bhavana Reddy Paluvai

Abstract: Diabetic retinopathy (DR) is a progressive retinal disease that can cause visual impairment and blindness. This study presents a five-class DR severity classification approach using retinal fundus images from the APTOS 2019 Blindness Detection dataset. The experimental workflow contains 3,662 labeled images representing five severity grades: no DR, mild, moderate, severe, and proliferative DR. Images are resized and processed using retinal-region cropping and Gaussian-based enhancement before being presented to transfer-learning models at 320?320 resolution. ImageNet-pretrained InceptionV3 and ResNet50 architectures are adapted for five-class classification. On the recorded experimental split, InceptionV3 achieved 78.3% test accuracy with a quadratic weighted Cohen's kappa of 0.822, while ResNet50 achieved 81.9% test accuracy with a kappa of 0.860. ResNet50 therefore produced the strongest overall result. The confusion matrix also reveals important class-specific limitations, particularly for the severe-DR class. The study provides a practical baseline for further improvement through stratified data splitting, training-only augmentation, consistent multiclass loss, class-aware learning, and external validation.

Keywords: Diabetic retinopathy, fundus images, medical image classification, transfer learning, ResNet50

How to Cite?: Tanuja Gannavarapu, Sai Prakash Gopalam, Bhavana Reddy Paluvai, "Transfer Learning for Five-Level Diabetic Retinopathy Severity Classification Using InceptionV3 and ResNet50", Volume 15 Issue 8, August 2026, International Journal of Science and Research (IJSR), Pages: 922-925, https://www.ijsr.net/getabstract.php?paperid=SR26812222953, DOI: https://dx.doi.org/10.21275/SR26812222953

Download Citation: APA | MLA | BibTeX | EndNote | RefMan