UC-Mamba: Adaptive Cross-Level Fusion Network With 2D-Selective Foveal Scanning and Channel-Gated Linear Unit for...

Accurate evaluation of ulcerative colitis severity remains challenging due to complex lesion characteristics and limitations in existing deep learning approaches. This paper presents UC-Mamba, a novel neural network architecture integrating biomimetic design principles with advanced feature extraction capabilities for enhanced UC assessment. To address critical challenges in current methodologies, we propose three innovative components: (1) a 2D-Selective Foveal Scanning mechanism to overcome the limited sensitivity to near-focal features and weaker perception of distant features in existing scanning methods; (2) a Channel-Gated Linear Unit to enhance local feature modeling, addressing the lack of local inductive bias in most existing networks; and (3) an Adaptive Cross-level Fusion mechanism to effectively integrate cross-layer features and capture rich global contextual information. Comprehensive experiments are conducted on two distinct UC datasets. On private Dataset, UC-Mamba achieves superior performance with 88.47 % accuracy, 84.63 % Matthews Correlation Coefficient, and 87.92 % F1-score, significantly outperforming state-of-the-art methods. Similar improvements are demonstrated on the more challenging public Dataset Kvasir, reaching 85.02 % accuracy. Ablation studies validate the effectiveness of each proposed component, with the complete model exhibiting optimal performance through their synergistic integration. The source code is available at github.com/kititoo/UC-Mamba.