UC-DiffuMambaNet: dual-scale state propagation and window attention for ordinal Mayo endoscopic grading

Automated four-grade Mayo Endoscopic Score (MES) classification is challenging because inflammatory severity depends not only on subtle local mucosal changes, but also on the collective spatial pattern of abnormalities that may be separated across the visible field. We present UC-DiffuMambaNet, a hybrid architecture that integrates dual-scale severity-evidence propagation, two-dimensional contextual coordination, and hierarchical ordinal prediction. The Severity Mamba Mixer first strengthens informative mucosal features through locally gated multi-scale enhancement and then processes them with parallel short- and long-retention selective state-space branches. This design preserves local feature continuity while linking separated severity cues over complementary contextual ranges. Medical Window Attention subsequently coordinates fine intra-window structure with communication across windows, complementing state propagation with explicit two-dimensional spatial context. Finally, an Auxiliary Hierarchical Classification Head refines ordered MES assessment through feedback-guided hierarchical prediction. Across three independent seeds, UC-DiffuMambaNet achieved ACC/QWK/Macro-F1 of 91.26%/96.28%/90.70% on Dataset A and 79.95%/87.65%/74.17% on Dataset B (LIMUC). Compared with the architecture-matched MambaVision-T baseline, ACC increased by 3.01 and 2.90 percentage points, respectively. Factorial and internal ablations confirmed complementary contributions from the three modules. Together, these components provide an integrated framework for ordinal MES assessment.