Explainable AI model improves early glaucoma detection

Muduli et al. publish an explainable CAD pipeline for early glaucoma detection in a bioRxiv preprint (Muduli et al., bioRxiv). Per the paper, the proposed system comprises four stages: image pre-processing, feature extraction, feature dimensionality reduction, and classification, and the authors report that their method reduces feature count while outperforming existing models in classification accuracy (Muduli et al., bioRxiv; Semantic Scholar). The paper specifically describes using an improved grey wolf optimization algorithm in tandem with an Extreme Learning Machine classifier.

Editorial analysis - technical context: The method reported by the authors pairs a metaheuristic feature selector, cited as IMGWO (improved grey wolf optimization), with ELM for classification. Combining metaheuristic selection with shallow, fast classifiers like ELM is a known strategy in medical-imaging tasks to lower dimensionality and speed inference, which can aid interpretability by shrinking the feature set to a smaller number of clinically meaningful inputs. The paper also emphasizes explainability through a reduced and interpretable feature set rather than through post-hoc attribution maps.

Editorial analysis: Glaucoma screening benefits from earlier, reliable detection because it often remains asymptomatic until significant vision loss. The paper frames its contribution as a potential second-opinion tool for ophthalmologists (Muduli et al., bioRxiv / Semantic Scholar). For practitioners, methods that lower feature dimensionality while maintaining or improving accuracy matter because they can reduce overfitting risk on small clinical datasets and simplify deployment in resource-constrained settings.

For practitioners: important next steps to assess the method's practical value include external validation on independent, demographically diverse retinal-fundus cohorts, reporting of sensitivity/specificity and ROC curves on held-out clinical data, and publication of code or model checkpoints to enable reproduction. Also watch for comparisons versus deep-learning baselines on identical datasets and for prospective or clinical-validation studies that quantify screening utility in real-world workflows.