MambaGLA-NPC: a 3D framework integrating CNN and state space models for robust multi-center nasopharyngeal carcinoma segmentation.
Authors
Affiliations (6)
Affiliations (6)
- School of Artificial Intelligence for Traditional Chinese Medicine, Zhejiang Chinese Medical University, Hangzhou, Zhejiang, China.
- Department of Radiology, Guangdong Provincial People's Hospital (Guangdong Academy of Medical Sciences), Southern Medical University, Guangzhou, Guangdong, China.
- Guangdong Provincial Key Laboratory of Artificial Intelligence in Medical Image Analysis and Application, Guangzhou, Guangdong, China.
- School of Pharmaceutical Sciences, Zhejiang Chinese Medical University, Hangzhou, Zhejiang, China.
- Department of Medical Imaging Center, Guangxi Medical University Cancer Hospital, Guangxi Clinical Research Center for Medical Imaging Construction, Guangxi Key Clinical Specialty (Medical Imaging Department), Dominant Cultivation Discipline of Guangxi Medical University Cancer Hospital (Medical Imaging Department), Nanning, China.
- School of Basic Medical Sciences, Zhejiang Chinese Medical University, Hangzhou, Zhejiang, China.
Abstract
Accurate 3D segmentation of nasopharyngeal carcinoma (NPC) from MRI is critical for radiotherapy planning but remains challenging due to poor multi-center generalization, lack of spatial continuity, and high computational costs of global modeling. We aim to address these limitations by proposing a novel 3D framework. We propose MambaGLA-NPC, a 3D deep learning framework integrating State Space Models (SSMs) and Convolutional Neural Networks (CNNs) in a U-shaped encoder-decoder architecture. A Conv-GLSSM module jointly extracts static local features via convolutions and global context via SSMs with a 3D local scanning strategy. A gated channel-spatial attention (GCS Att) module suppresses irrelevant noise in skip connections, and an NPC-specific adapter enables parameter-efficient fine-tuning across multi-center datasets. On a multi-center dataset of 1,098 NPC patients from three hospitals, our model achieves state-of-the-art performance on the internal test set (Dice: 80.47%, IoU: 67.83%, HD95: 4.09 mm). Under cross-center evaluation, it significantly outperforms prior methods (e.g., Dice of 87.13% and 76.70% on two external centers). The NPC-Adapter boosts cross-center Dice by up to 8.84% compared to no adaptation. Ablation studies confirm the contribution of each module, and the Conv-GLSSM reduces computational cost versus Transformers (21.66 GFLOPs vs. OOM for self-attention at <math xmlns="http://www.w3.org/1998/Math/MathML"> <mrow> <msup><mrow><mn>128</mn></mrow> <mrow><mn>3</mn></mrow> </msup> </mrow> </math> resolution). MambaGLA-NPC is the first Mamba-based 3D NPC segmentation framework that effectively balances local-global feature integration with linear complexity, achieving superior accuracy and generalization across multi-center MRI data while maintaining high efficiency.