Robust Glaucoma Prediction from Fundus Images using DenseNet201 &NASNetMobile

Electronics and Communication Engineering

Glaucoma is a progressive eye disease affecting approximately 64 million people globally, leading to damage of the optic nerve head (ONH) and potential irreversible blindness. Early detection is critical to prevent vision loss; however, traditional clinical approaches, such as manual segmentation of the optic cup and disc for cup-to-disc ratio (CDR) calculation, are often time-consuming, subjective, and dependent on expert evaluation. To overcome these limitations, this study proposes an automated glaucoma detection system based on deep learning, utilizing retinal fundus images for binary classification of healthy and glaucomatous eyes. The system employs two advanced convolutional neural networks, DenseNet201 and NASNetMobile, both trained using transfer learning techniques and fine-tuned for optimal performance. Preprocessing techniques such as histogram equalization are applied to enhance image contrast, and class imbalance is managed using computed class weights. The model focuses on extracting key features from the optic disc (OD) and optic cup (OC) areas, emphasizing crucial indicators like the cup-to-disc ratio (CDR) and the neuroretinal rim (NRR) structure. Performance evaluation is conducted by comparing the classification accuracy, precision, and recall of both models to identify the more effective solution for practical glaucoma screening. DenseNet201, recognized for its deep feature extraction capabilities, and NASNetMobile, optimized for lightweight deployment, offer valuable insights into achieving a balance between accuracy and computational efficiency. This research highlights the promising role of deep learning in supporting clinicians with faster, more objective, and reliable glaucoma diagnosis.

This work is licensed under a Creative Commons Attribution-ShareAlike 4.0 International License.