Precision Glaucoma Screening via BWO-Guided Instance Segmentation in an SRResNet-RCNN Framework

Abstract

Glaucoma is still one of the most common, irreversible cause of blindness in the world with the worldwide patient burden estimated to be over 110 million by 2040. Automated and objective screening tools are clinically essential as structural deterioration of the optic nerve is asymptomatic in early disease. In this article a novel three-stage computational pipeline, named BWO-RCNN, is introduced which combines bio-inspired parameters optimization with deep-learning segmentation for high-accuracy glaucoma detection of retinal fundus photographs. In the initial stage, a hybrid pre-processing module using Median Filtering and Contrast Limited Adaptive Histogram Equalization (CLAHE) is used to reduce noise and retain the edge detail that is important for diagnosis. The second stage is to optimise the downstream Region-based Convolutional Neural Network (RCNN) using a Black Widow Optimisation (BWO) algorithm to learn the best hyper-parameter settings. The third stage is that the SRResNet architecture is employed as a backbone for feature extraction, and RCNN is applied for joint instance segmentation of the optic-disc (OD) and the optic-cup (OC). Evaluation of the publicly available Kaggle dataset of 650 fundus images shows that BWO-RCNN has a classification accuracy of 97 %, precision of 98 %, recall of 95 % and an F1-score of 96 %, outperforming the benchmark methods CNN, RF-DNN and FCN in all cases. The results highlight the importance of the synergy between swarm-intelligence optimisation and residual super-resolution networks in the field of clinical image analysis.