Neurosymbolic AI Framework for Explainable Retinal Disease Classification From OCT Images

Purpose: Accurate classification of retinal diseases such as dry age-related macular degeneration, wet AMD, epiretinal membrane, full-thickness macular hole (MH), lamellar MH, and central serous chorioretinopathy (CSC) is essential for effective treatment and clinical decision-making. Traditional deep learning models, however, often struggle with imbalanced datasets and lack interpretability, limiting their translational applicability in ophthalmology. Methods: We propose a neurosymbolic framework that integrates a convolutional neural network (CNN) with a symbolic reasoning layer based on expert-defined clinical rules. A total of 10,846 optical coherence tomography images were retrospectively collected and categorized into seven diagnostic classes: dry AMD, wet AMD, epiretinal membrane, full-thickness MH, lamellar MH, central serous chorioretinopathy, and healthy retinas. Results: Our neurosymbolic model achieved macro-precision 0.83, recall 0.82, and F1 0.81, on internal dataset, having slightly better performance than the CNN (0.64/0.83/0.68). On the external dataset, it retained superior performance, macroprecision 0.85, recall 0.79, F1 0.78, versus the CNN (0.73/0.64/0.59). Conclusions: Our hybrid neurosymbolic framework introduces a unified paradigm that couples symbolic reasoning with a conventional CNN, improving diagnostic performance while delivering transparent, clinically interpretable decisions. It is particularly effective for rare and complex conditions that often challenge end-to-end deep learning models. Translational Relevance: By integrating symbolic clinical logic with visual pattern recognition, the neurosymbolic model fosters trust in artificial intelligence–assisted diagnostics and supports precise, explainable decision-making in retinal care.