Open Access

ARTICLE

A Novel Approach Deep Learning Framework for Automatic Detection of Diseases in Retinal Fundus Images

Kachi Anvesh^1,2, Bharati M. Reshmi^2,3, Shanmugasundaram Hariharan⁴, H. Venkateshwara Reddy⁵, Murugaperumal Krishnamoorthy⁶, Vinay Kukreja⁷, Shih-Yu Chen^8,9,*

1 Department of Information Technology, Vardhaman College of Engineering, Shamshabad, Hyderabad, 501218, India
2 Department of Computer Science and Engineering, Visvesvaraya Technological University, Belagavi, 590018, India
3 Department of Artificial Intelligence and Machine Learning, Basaveshwar Engineering College, Bagalkote, 587102, India
4 Department of Artificial Intelligence and Data Science, Vardhaman College of Engineering, Hyderabad, 501218, India
5 Department of Computer Science and Engineering, Vardhaman College of Engineering, Hyderabad, 501218, India
6 Department of Electrical and Electronics Engineering, Vardhaman College of Engineering, Hyderabad, 501218, India
7 Centre for Research Impact & Outcome, Chitkara University Institute of Engineering and Technology, Chitkara University, Punjab, 140401, India
8 Department of Computer Science and Information Engineering, National Yunlin University of Science and Technology, Yunlin, 64002, Taiwan
9 Intelligence Recognition Industry Service Research Center, National Yunlin University of Science and Technology, Yunlin, 64002, Taiwan

* Corresponding Author: Shih-Yu Chen. Email:

Computer Modeling in Engineering & Sciences 2025, 143(2), 1485-1517. https://doi.org/10.32604/cmes.2025.063239

Received 09 January 2025; Accepted 27 March 2025; Issue published 30 May 2025

Abstract

Automated classification of retinal fundus images is essential for identifying eye diseases, though there is earlier research on applying deep learning models designed especially for detecting tessellation in retinal fundus images. This study classifies 4 classes of retinal fundus images with 3 diseased fundus images and 1 normal fundus image, by creating a refined VGG16 model to categorize fundus pictures into tessellated, normal, myopia, and choroidal neovascularization groups. The approach utilizes a VGG16 architecture that has been altered with unique fully connected layers and regularization using dropouts, along with data augmentation techniques (rotation, flip, and rescale) on a dataset of 302 photos. Training involves class weighting and critical callbacks (early halting, learning rate reduction, checkpointing) to maximize performance. Gains in accuracy (93.42% training, 77.5% validation) and improved class-specific F1 scores are attained. Grad-CAM’s Explainable AI (XAI) highlights areas of the images that are important for each categorization, making it interpretable for better understanding of medical experts. These results highlight the model’s potential as a helpful diagnostic tool in ophthalmology, providing a clear and practical method for the early identification and categorization of retinal disorders, especially in cases such as tessellated fundus images.

---

Keywords

---