Abstract
Visual impairment affects over 2.2 billion people globally and results in an estimated 3 trillion in economic loss, with retinal diseases such as Choroidal Neovascularization (CNV), Diabetic Macular Edema (DME), and Drusen contributing significantly. Optical Coherence Tomography (OCT) is the standard for retinal imaging, however its diagnostic potential is hindered by speckle: an interference pattern caused by coherent light scattering within heterogeneous retinal tissue. Most existing deep learning models treat speckle as noise and suppress it, often at the cost of losing biologically relevant information. To address this, we propose a Speckle-Aware Dynamic Vision Transformer (SA-DVT), a novel deep learning model that leverages speckle as a diagnostic feature rather than suppressing it. The model integrates a robust speckle map generator with a learnable gating mechanism that dynamically blends structural and speckle content. This enhanced representation is then passed to a modified Swin Transformer backbone adapted for single-channel OCT inputs, enabling classification of CNV, DME, Drusen, and Normal scans. The SA-DVT was evaluated across seven public OCT datasets and a balanced 60,000-image split. It achieved over 96% accuracy with high per-class precision and recall. These results demonstrate strong generalizability across varied imaging conditions and diseases. SA-DVT offers a biologically informed and clinically robust solution for retinal disease classification from a single OCT scan by treating speckle as signal rather than noise.