Uncertainty-Driven Global-Local Mamba with UNet for gastric cancer histopathological image classification

Histopathological image classification is important for gastric cancer diagnosis. Existing methods have difficulties in balancing global context modeling and local texture extraction. A novel Uncertainty-Driven Global-Local Mamba with UNet to proposed to solve these problems. It is used for gastric cancer histopathological image classification. The new model has three key innovations. First it has an uncertainty-driven scanning mechanism. This mechanism adjusts the attention weights of global and local feature extraction dynamically. It is based on the confidence of intermediate features and can highlight suspicious lesion regions effectively. It can also suppress irrelevant background noise. Second it has a global-local Mamba module. This module combines bidirectional selective scanning and depthwise separable convolution. Bidirectional selective scanning is used for long-range dependency modeling. Depthwise separable convolution is used for local texture enhancement. Compared with transformer-based methods it achieves linear computational complexity. Third it has UNet-inspired skip connections. These connections fuse multi-scale features from the encoder and decoder. Extensive experiments are conducted on the GasHisSDB dataset. The new approach achieves excellent performance. Its accuracy is 98.76%. Its precision is 98.53%. Its recall is 98.31%. Its F1-score is 98.42%. The new method provides a reliable and efficient tool for computer-aided diagnosis of gastric cancer. It has potential for clinical deployment in pathological analysis workflows.