PhysioSpeck-Net: Physics-Guided Feature Separation for Parameter-Efficient and Uncertainty-Aware Retinal OCT Classification Under Cross-Dataset Shift

Abstract

Deep networks for optical coherence tomography (OCT) classification usually learn directly from image appearance, which leaves coherent speckle, depth-dependent attenuation, point-spread-function blur, and refractive distortion entangled with pathology inside a single feature representation. This work asks whether making that separation architecturally explicit is a useful inductive bias, rather than whether it raises benchmark accuracy. PhysioSpeck-Net routes a shared stem representation through four branches aligned one-to-one with distinct OCT image-formation effects, recombines them through cross-branch attention and a physics-guided gate, and constrains them with auxiliary objectives derived from a seven-layer retinal simulator. The framework is evaluated on a 1000-image balanced test set and, without fine-tuning, on 1400 images from a second OCT source. Under a common training protocol, the model is competitive with substantially larger convolutional and transformer baselines while using 8.80 million parameters, but the more informative results concern reliability: predictive entropy separates errors from correct predictions with error-detection AUCs of 0.9847 and 0.9688, expected calibration error remains near 4–5% on both sets, and Grad-CAM++ perturbation analysis shows that attribution faithfulness is class-dependent rather than uniformly high. Distributional analysis further shows a residual gap between simulated and clinical images. Leave-one-branch-out and loss ablations produced consistent performance reductions, paired significance testing confirmed gains over most conventional baselines, and duplicate-control analysis found no exact or confirmed near-duplicate images across the evaluated datasets. These findings support physics-guided feature separation as a compact and uncertainty-aware strategy for retinal OCT classification under dataset shift.