A framework for robust glaucoma detection: A confidence-aware deep uncertainty quantification approach with a comprehensive assessment for enhanced clinical decision-making

Glaucoma poses a significant threat to public health worldwide, as it can result in irreversible vision loss. Timely identification is vital for halting the progression of visual field deterioration. In recent years, deep neural networks (DNNs) have become increasingly popular in medical imaging due to their ability to identify patterns. As a result, this study introduces a new computer-aided diagnosis (CAD) system based on deep learning (DL) algorithms for glaucoma detection that extracts meaningful features from retinal fundus images (RFIs) and employs uncertainty quantification (UQ) models, including Monte Carlo dropout (MCD), ensemble Bayesian, and ensemble Monte Carlo dropout (EMCD), to generate both point estimates and confidence values for the outputs, thereby capturing the uncertainty associated with the classifications. The proposed framework is validated using well-known clinical datasets, and the reliability of the outputs is evaluated using comprehensive performance metrics such as expected calibration error (ECE), entropy analysis, and a multi-criteria UQ assessment. Experimental results demonstrate the superiority of the ensemble model, with uncertainty accuracies registering at 97.64%, 97.26%, and 98.97% for the “ACRIMA”, “RIM-ONE-DL”, and “ORIGA” datasets, respectively. Moreover, the proposed algorithms can alert users to the majority of erroneous diagnoses by assigning uncertainty labels, providing valuable insights for clinicians in glaucoma detection. Such tools can assist healthcare professionals in reducing the probability of misdiagnosis and ensuring that patients receive timely and appropriate treatment.