Spatially Grounded Neurosymbolic AI for Explainable Retinal Disease Classification From OCT Images

Abstract

Purpose: To develop and evaluate a spatially grounded neurosymbolic AI framework for explainable multilabel retinal disease classification from optical coherence tomography (OCT) images, integrating convolutional neural network (CNN)–based retinal sign extraction, symbolic clinical reasoning, and retinal-layer spatial information. Methods: Spectral-domain OCT B-scans were retrospectively collected and classified into ten categories: dry and wet age-related macular degeneration, central serous chorioretinopathy, myopic chorioretinopathy, epiretinal membrane, pseudomacular hole, lamellar macular hole, full-thickness macular hole, diabetic retinopathy, and normal retina. The model combined a CNN backbone with SegFormer-derived retinal-layer maps and mask-guided attention to estimate continuous OCT sign probabilities, which were propagated through expert-defined symbolic rules. Performance was evaluated using multilabel precision, recall, F1-score, confusion matrices, and receiver operating characteristic analysis. Inference was also performed on 402 OCT images from pathologies outside the target taxonomy. Results: Micro-average precision, recall, and F1-score were 0.91, 0.81, and 0.86, respectively; weighted-average values were 0.91, 0.81, and 0.85, and macro-average values were 0.78, 0.71, and 0.74. The micro-average AUC was 0.98. None of the 402 out-of-taxonomy pathological OCT images received a positive Normal-class prediction at the predefined 0.5 threshold. Conclusions: The spatially grounded neurosymbolic framework provided interpretable multilabel OCT classification and showed conservative behavior on pathological images outside the modeled taxonomy. This should not be interpreted as diagnostic recognition of unseen diseases. Translational Relevance: The framework may support transparent, sign-based, and anatomically grounded AI-assisted OCT interpretation.