Key points are not available for this paper at this time.
Glaucoma, a leading cause of irreversible blindness, necessitates accurate and early diagnosis to prevent visual impairment. Current methods relying solely on OCT A-scans often fall short in specificity and sensitivity, and may overlook subtle anomalies indicative of early glaucoma. Furthermore, delays in diagnosis can contribute to disease progression. This paper introduces an innovative deep learning model that amalgamates OCT A-scans, fundus images, and visual field data, leveraging the strengths of multi-modal fusion to provide a comprehensive analysis for glaucoma diagnosis. Transfer learning, utilizing pre-trained models like InceptionV3 and ResNet50, was employed to enhance feature extraction, alongside anomaly detection methods including One-Class SVM and Isolation Forest to identify atypical glaucoma presentations. Moreover, the integration of an attention mechanism in segmentation models, specifically V-Net and SegNet, facilitated precise delineation of regions of interest in OCT A-scans. The proposed model significantly outperformed existing methods, enhancing precision by 8.5%, accuracy by 3.9%, recall by 8.3%, AUC by 4.9%, and specificity by 5.5%, while concurrently reducing diagnosis delay by 10.4%. These substantial improvements underscore the potential of our method in revolutionizing glaucoma diagnosis, ultimately contributing to timely intervention and better patient outcomes.
Mundada et al. (Tue,) studied this question.