Fused-AETNet: A Variational Transformer-Based Framework for Diabetic Retinopathy Classification Using OCT Biomarkers

Early detection and accurate diagnosis of Diabetic Retinopathy (DR) using Optical Coherence Tomography (OCT) images remain challenging due to subtle pathological alterations across multiple retinal layers. In this paper, we present Fused-AETNet (Fused AutoEncoder–Transformer Network), a novel deep-learning framework that uniquely combines clinically interpretable handcrafted OCT biomarkers with a robust variational autoencoder and a transformer-based classifier. Unlike prior OCT–DR approaches that directly utilize raw intensities or generic CNN-derived features, our method begins with precise segmentation of eleven retinal layers in the central OCT B-scan using an atlas-guided probabilistic approach. From each layer, we extract three clinically validated morphological descriptors—reflectivity, Laplacian-based thickness, and boundary tortuosity—and summarize them through percentile-based Cumulative Distribution Functions (CDFs) to capture regional variability in a compact, size-invariant form. These descriptors are fused into a smooth latent embedding via variational autoencoding, enabling robust modelling of cross-layer dependencies and uncertainty. The latent space is then refined using a lightweight transformer-inspired attention module that selectively amplifies diagnostically relevant embeddings while suppressing irrelevant variations. This biomarker-driven design enhances both the specificity and interpretability of the diagnostic decision process. Our training employs a composite loss function that jointly enforces accurate descriptor reconstruction, latent space regularization, and high classification accuracy. We validated the framework on OCT scans from 481 subjects, achieving superior performance compared to traditional machine learning classifiers and state-of-the-art deep networks, with an average accuracy of $93.08\% \pm 4.21$ , precision of $93.33\% \pm 4.84$ , recall of $96.00\% \pm 5.96$ , and F1 score of $94.48\% \pm 3.21$ . By explicitly modelling layer-wise biomarkers and integrating attention-guided latent feature selection, Fused-AETNet offers a compact, interpretable, and clinically relevant solution for OCT-based DR diagnosis.