DenseNet-169 and PatchTST-based Early Glaucoma Detection using Retinal Fundus Images

Glaucoma is a chronic ophthalmic disease that is characterized by progressive optic nerve damage. If untreated, it leads to permanent vision loss. Hence, an automated detection model is essential to detect glaucoma at an early stage to prevent from blindness. However, the conventional Deep Learning (DL) models struggle to learn the various characteristics that consist of Optic Disc (OD) with different sizes, shapes and positions, which results in detecting the early stage of glaucoma as normal. To solve this problem, in this research, a Patch Time Series Transformer (PatchTST) is proposed to classify between normal and glaucoma effectively. For glaucoma detection, the raw retinal fundus images from the two benchmark datasets are acquired and preprocessed using standard techniques to normalize image quality. Subsequently, a pretrained DenseNet-169 architecture is used for deep feature extraction; its dense connectivity pattern continuously concatenates multi-scale feature maps to capture both macro-level boundaries for varied OD sizes and micro-level localized textures. The PatchTST method maps the extracted deep features into tokenized sequence patches to learn fine-grained disease characteristics to classify healthy and diseased eyes efficiently. Experimental results demonstrate that the proposed method achieves classification accuracy of 98.84% on DRISHTI-GS and 98.98% on RIM-ONE datasets, respectively, while representing competitive glaucoma detection performance compared with existing methods, including Attention U-Net, with a modified ResNet-50 (AU-ResNet-50), under identical experimental conditions.