<![CDATA[Glaucoma is a leading cause of irreversible vision loss, and its early detection remains challenging due to the presence of subtle structural varitions in retinal fundus images. Convolutional neural networks (CNNs) have demonstrated strong performance for automated classification of glaucoma; however, the relationship between extracted features and prediction outcomes frequently proves challenging to interpret. Most existing explainable artificial intelligence (XAI) approaches rely on post hoc visualizations, which provide limited insight into the decisions-making process. To address this limitation, this present study proposes a hybrid CNN–feature selection–ANFIS framework (CNN–FS–ANFIS) that integrates interpretability directly within the decision layer. In this framework, the first stage of the process involves adapting a CNN backbone to the glaucoma classification task through the use of transfer learning. This is then used as a fixed feature extractor to obtain retinal representations for decision-layer modeling. Subsequently, a feature selection stage is applied driven by sparsity to construct a compact and structured subset of informative features. These features are then fed into an Adaptive Neuro-Fuzzy Inference System (ANFIS), enabling predictions to be expressed through explicit fuzzy rule-based reasoning. The impact of feature compactness is examined in a controlled experimental setting, where the feature subset size is varied from three to nine. The findings demonstrate that compact feature subsets can achieve consistent and competitive performance. By means of LASSO-selected features, the ANFIS decision layer achieved an AUC of 0.84±0.01, sensitivity of 0.82±0.13, specificity of 0.74±0.10, and an F1-score of 0.79±0.04. Rule-base analysis further exhibited that two-to three-rule ANFIS configurations-maintained AUC values of approximately 0.84 while preserving a transparent and manageable decision structure. The proposed framework, therefore, enables direct analysis of the relationship between selected CNN features, fuzzy rules, and model outputs. This traceable decision pathway has the potential to support more transparent and auditable glaucoma screening systems.]]>
