Paying Attention to the Input Layer: Class-Conditioned Attribution with Kolmogorov-Arnold Patch Embeddings for OCT

Deep learning performs well in retinal optical coherence tomography (OCT) classification, but its internal operations remain difficult to inspect. This motivates post hoc attribution and architectures with readable components, including Kolmogorov-Arnold Networks (KANs), whose learned univariate functions can be plotted directly. Existing vision applications commonly place KAN components in classifier heads or intermediate blocks, where their functions describe learned features rather than input processing. We place a KAN-based spline bank at a Vision Transformer’s patch embedding to inspect learned transformations of normalized OCT intensity. The nonlinear intensity remapping remains exactly decomposable through the linear projection into tokens. Combining this decomposition with attention rollout yields Spline-Exact Rollout Attribution (SERA). We show that the previously proposed Interpretable Capacity (IC) metric is class-invariant: all 60 pairwise class-label comparisons across 10 OCT-17 images produced identical IC maps. We therefore propose Faithful Attention Shift (FAS), replacing IC’s evidence with class-conditioned SERA. SERA maps changed in all 60 comparisons, assigned 1–3% of attribution mass to corners covering 9% of image area versus 48–53% for Eigen-CAM, and yielded larger top-10% deletion confidence drops (0.45–0.71 versus −0.004 to 0.14). Balanced accuracy remained 0.963 on OCT-17 and 0.917 on OCTID.