Simplex-Aligned Diffusion with Cross-Granularity Interaction for Robust Medical Image Classification

The clinical deployment of medical image classification systems hinges on their trustworthiness, specifically, the ability to provide calibrated uncertainty estimates and maintain robustness under acquisition shifts. While generative diffusion models offer promising dis-tributional modeling, existing approaches suffer from a fundamental geometric conflict: they apply unbounded Gaussian noise directly to bounded label simplices. We identify that this theoretical mismatch forces predictions into invalid probability spaces, serving as a primary source of model unreliability and overconfidence. To resolve this, we propose Simplex-Aligned Diffusion. Unlike standard methods, we reformulate the label generation process on an unconstrained logit manifold. By mapping the probability simplex to a Euclidean space, we ensure mathematical consistency with Gaussian diffusion, which effectively acts as a geometric regularizer for uncertainty calibration. Furthermore, we introduce a Transformer-based Cross-Granularity Interaction module to stabilize visual guidance by dynamically modeling global-local dependencies. Extensive experiments on the APTOS2019 and HAM10000 benchmarks demonstrate that our framework not only achieves competitive accuracy but significantly outperforms state-of-the-art baselines in calibration error (ECE) and resilience to clinical artifacts (e.g., sensor noise, blur), offering a mathematically rigorous and clinically reliable paradigm.¹