GTMamba: Graph Tri‐Orientated Mamba Network for 3D Brain Tumor Segmentation

ABSTRACT Recently, Mamba has garnered increasing attention due to its efficiency and effectiveness in modeling long‐range dependencies. However, adapting it to non‐sequential brain tumor image data remains a significant challenge. To address this, we propose the Graph Tri‐orientated Mamba network (GTMamba) for brain tumor image segmentation. This network is capable of flexibly capturing the relationships between vertices and their neighboring vertices, thereby enhancing the selection mechanism of the Mamba module. This improvement allows the network to better adapt to non‐sequential image data and significantly enhances segmentation accuracy. On the BraTS 2021 and MSD Task01_BrainTumour datasets, GTMamba achieved Dice values of 94.29%/92.08%, 94.01%/90.58%, and 88.44%/74.02% for the whole tumor, tumor core, and enhanced tumor segmentation tasks, respectively. Compared to other state‐of‐the‐art methods, GTMamba demonstrates superior overall performance in terms of segmentation accuracy and parameter efficiency.