Glaucoma detection from retinal fundus images using graph convolution based multi-task model

Abstract

• Pioneering Multi-task Learning Approach: Our research introduces a cutting-edge multi-task deep learning model tailored for automated segmentation and classification in glaucoma detection, leveraging retinal fundus images. This represents a significant leap in the application of artificial intelligence within ophthalmology. • Exceptional Diagnostic Accuracy: The model showcases unprecedented diagnostic performance, achieving up to 97.43 % accuracy and an AUROC of 0.985 for glaucoma detection. Such high levels of accuracy promise a transformative impact on early glaucoma screening practices. • Innovative Integration of Technologies: By incorporating Mobile-Netv2 in the encoder, Graph Convolution Network (GCN) in the decoder, and an attention module for precise feature extraction, our model epitomizes the forefront of innovation in deep learning for medical imaging. • Robust Validation Across Diverse Datasets: The effectiveness and versatility of our model are thoroughly validated across multiple fundus image datasets, including ORIGA, REFUGE, and DRISTI-GS, underscoring its robustness and applicability in varied clinical settings. • Contributions to Preventive Healthcare: Given the model's potential to significantly enhance early detection and intervention strategies for glaucoma, our work stands as a beacon for the future of preventive healthcare, potentially mitigating the risk of irreversible vision loss for millions worldwide. Glaucoma is an abnormality in the eye condition that, if not treated within a safe time limit, can result in visual loss. Glaucoma diagnosis requires professionals to identify minor structural changes in the structure of the optic disc and optic cup from retinal fundus images in a short period. Deep learning algorithms have been employed effectively in the segmentation of biomedical images and the identification of diseases. To accomplish good generalization, model training requires comprehensive annotations, which is a difficult task. The intended objective of the present study is to come up with and train a distinctive multi-task deep learning model for automated fundus image segmentation and classification. The multi-task model learns for the segmentation task of Optic Disc (OD) and Optic Cup (OC) and the classification task for accurate glaucoma detection using both structural and image-based features. The multi-task model proposed a modified U-net architecture in which Mobile-Netv2 is used in the encoder part, Graph Convolution Network (GCN) is used in the decoder part, and an attention module (AM) is used to locate the region of interest (ROI) for better feature extraction. The implementation of this model is done using three fundus image datasets such as ORIGA, REFUGE, and DRISTI-GS. The performance of the proposed multi-task model is compared with some existing methods. It shows maximum accuracy of 97 . 43 % and AUROC of 0.985 for the glaucoma detection task and high-quality OD and OC segmented images with dice coefficient of 97 . 95 % and 96 . 11 % respectively for the segmentation task.

Topics & Concepts