Integrated deep learning model for multi-label retinal disease diagnosis.
Authors
Affiliations (2)
Affiliations (2)
- Department of Information Systems, College of Computer and Information Sciences, Jouf University, Sakaka, 72314, Kingdom of Saudi Arabia. [email protected].
- Department of Information Systems, College of Computer and Information Sciences, Jouf University, Sakaka, 72314, Kingdom of Saudi Arabia.
Abstract
The multi-label retinal disease classification is a difficult one since fundus images might comprise many co-occurring abnormalities, extreme in relation to the class, as well as significant lesion appearance variation. In order to deal with this issue, this paper suggests a combined deep learning architecture to diagnose retinal diseases automatically with the help of the MuReD dataset, which includes 2208 fundus images that were gathered at ARIA, STARE, and RFMiD and reformulated with 19 retinal disease labels, excluding the non-diagnostic category of OTHERS. The proposed model is based on a hybrid convolutional structure that integrates two complementary branches of feature extraction and implements a multi-stage preprocessing pipeline consisting of contrast enhancement, luminance correction, noise reduction, and retinal masking. Fusion and optimization of extracted feature representations are performed to perform multi-label prediction with binary supervision of label-wise predictions as sigmoid outputs. In order to assess the value of architecture and preprocessing design, ablation experiments were performed on a great scale by using single-backbone baselines, hybrid configurations, and preprocessing-specific configurations. According to the results, the highest F1-micro = 0.5484, PR-AUC micro = 0.5696, and ROC-AUC micro = 0.9349 were observed in the best variant of the hybrid, and the best PR-AUC micro = 0.5821 and ROC-AUC micro = 0.9369 were observed in another hybrid variant. Other tests on bootstrap confidence interval, calibration evaluation, Monte Carlo dropout, and repeated multi-seed experiments also indicated the strength of the proposed model. These results reveal that feature fusion with built-in deep learning is an effective solution to multi-label retinal disease diagnosis and has a potential real-world application to computer-aided screening and triage in ophthalmology.