MamTrans: magnetic resonance imaging segmentation algorithm for high-grade gliomas and brain meningiomas integrating attention mechanisms and state-space models.

Background: Meningiomas and gliomas represent the most common benign and malignant brain tumors, where accurate segmentation is essential for clinical assessment and surgical planning. Although magnetic resonance imaging (MRI) serves as a crucial diagnostic tool, precise segmentation remains challenging due to significant morphological and structural variations between tumor types and surrounding complex soft tissues. While Mamba models demonstrate excellence in sequence processing and attention mechanisms show promising performance, both face limitations in feature extraction and computational efficiency, respectively. To address these challenges, we propose the MamTrans algorithm, which integrates state-space models (SSMs) with attention mechanisms to significantly improve computational efficiency while maintaining segmentation accuracy.

Methods: This study utilized 418 cases of axial T1-weighted contrast-enhanced MRI data of brain tumors, comprising 177 cases of high-grade gliomas and 241 cases of meningiomas. To validate the findings, five-fold cross-validation was employed.

Results: The newly algorithm MamTrans achieved promising segmentation results in the high-grade glioma segmentation experiment, with an intersection over union (IoU) of 88.12, a Dice similarity coefficient (DSC) of 89.23, and a Hausdorff distance (HD) of 12.67. In the meningioma segmentation experiment, its segmentation metrics were IoU of 90.26, DSC of 91.27, and HD of 15.14, on the external dataset, the model obtained IoU of 90.34, DSC of 91.25, and HD of 14.17, outperforming other segmentation models such as U-Net, DeepLab, and Attention U-Net.

Conclusions: The research results demonstrate that the proposed MamTrans algorithm outperforms various segmentation models in the segmentation tasks of gliomas and meningiomas. Innovatively, this single algorithm achieves high-precision segmentation for two tumor types with remarkably different morphologies, while significantly reducing model complexity and computational overhead, exhibiting substantial clinical application value.