1. Deep learning-Assisted Glaucoma Diagnosis and Model Design
- Author
-
J. Surendiran and M. Meena
- Subjects
characteristic ,diagnosis ,TVGH ,Glaucoma ,GON ,enormous ,CNN ,classifier - Abstract
Purpose: To evaluate the key factors used for classification and methods to improve the glaucomatous optic neuropathy (GON) in the form of systematic study along with fundus pictures as constraint sample be detected. Methods: Retrospectively, 940 fundus photographs from Eye hospital were gathered. The clinical and demographic details were noted together with the constitutional and functioning metrics of the pictures with GON. Convectional neural networks (CNNs) were built using transfer learning based on VGGNet to recognize GON. When CNN classifier had least rating the concluding classification will be construct with the cup-to-disc would be created by a extractor named support vector Machine. Scores were used to prevent missing instances with advanced GON. The TVGH dataset was used to create the CNN classifier, which was subsequently improved by fusing the training pictures from the TVGH and Drishti-GS datasets. CNN's primary characteristics for classification were identified using the class activation map (CAM). Classifier’s performance were evaluated with Area under receiver operating characteristic curve (AUC) and the diagnostic accuracy will enable to ccomparing the ensemble model. Results: While the comparison to ensemble model's accuracy rate of 92.8 percent, the CNN classifier's accuracy, sensitivity, and specificity were 95.0 percent, 95.7 percent, and 94.2 percent, respectively, in 187 EH test images, respectively, in 187 EH test pictures, and the AUC was 0.992 as opposed to the ensemble model's accuracy rate of 92.8 percent. The accuracy of the CNN, the fine-tuned CNN, and the ensemble model for the Drishti-GS test pictures was 33.3 percent, 80.3 percent [33], and 80.3 percent, respectively. Neither moderately nor severely ill photos were incorrectly classified using the CNN classifier. Class-discriminate areas discovered by CAM co-localized with established GON traits. Conclusions: When enormous image datasets are not readily accessible for developing evident learning model towards glaucoma diagnosis, The combo model or a personalized CNN classifier might be as practicable designs.
- Published
- 2023