Abstract
The prevalence of retinal diseases demonstrates the need for more efficient diagnostic methods. Conditions such as choroidal neovascularization, diabetic macular edema, and drusen remain leading causes of vision loss. Optical Coherence Tomography (OCT) is the clinical gold standard for retinal imaging, yet manual interpretation is time consuming and subject to variability across physicians. Clinical Decision Support Systems (CDSS) address these limitations by aiding disease detection; however, current CDSSs often lack transparency, limiting clinical adoption. Interpretability must encompass graphical, visual, and textual modalities that mirror physician reasoning. This study introduces a unified interpretable framework integrating all three. t-Distributed Stochastic Neighbor Embeddings generates projections of learned embeddings, revealing class separation and progression trends, that provide graphical interpretability. Swin-UNet generates a retinal layer segmentation image that provides visual interpretability by supporting anatomical verification. GPT-5 generates a structured, evidence-grounded patient report that provides textual interpretability. Evaluations on benchmark OCT datasets confirm consistent cluster geometry aligned with severity, accurate layer delineation, and verifiable narrative outputs, advancing explainable and workflow-ready CDSS for retinal disease management.