Increasing the speed of diagnosis of glaucoma by using multitask deep neural network from retinal images

Glaucoma stands out as a prevalent ocular ailment in the elderly population, causing substantial harm to the optic nerves and eventual vision impairment. Fundus photography plays a pivotal role in the clinical assessment of glaucoma, facilitating the exploration of associated morphological alterations. Computational algorithms, capable of processing fundus images, have emerged as indispensable tools in this diagnostic domain. Hence, the imperative development of an automated diagnostic system leveraging image processing techniques is underscored. In this study, a novel approach to the segmentation and classification of retinal optic nerve head images is introduced. This method concurrently executes both tasks through a deep learning framework, thereby enhancing the learning speed within the network. The proposed network encompasses approximately 29 million parameters and demonstrates an efficiency of 2.5 seconds for segmenting and classifying retinal images. Central to this strategy is a multi-task deep learning network, harmonizing segmentation and classification processes, and leveraging information from both tasks to optimize learning efficacy. Validation of the proposed method is conducted using the publicly available ORIGA dataset. The attained performance metrics for accuracy, sensitivity, specificity, and F1-score are 99.461, 93.46, 100, and 98.7006, respectively. These results collectively affirm the substantial advancement achieved by the proposed method in comparison to existing methodologies.