基于多模态磁共振成像与尺度特征差异的胶质瘤分割算法

打开文本图片集

中图分类号：TP312;R739.41 文献标志码：A

DOI：10.13338/j.issn.1674-649x.2025.03.015

Glioma segmentation algorithm based on multimodal magnetic resonance imaging and scale feature differences

WEI Wei¹，LYU Xin1'²，MA Menghang1， HU Zhenyuan 1 ，LIU Xiangyu³ ，LIU Qinfeng4 ，LIAO Guisheng

（1.School of Electronics and Information，Xi'an Polytechnic University，Xi'an 71oo48，China;

2.School of Information Engineering，Chang'an University，Xi'an 71oo64，China; 3.School of Life Science and Technology，Xidian University，Xi'an 7lOl26，China; 4.Medical Equipment Management Department， Shaanxi Provincial People's Hospital， Xi'an 710068，China; 5.National Key Laboratory of Radar Signal Processing， Xidian University， Xi'an 7lO071，China）

Abstract To address the challenges of tumor boundary delineation in multimodal magnetic resonance imaging caused by uneven data distribution and ambiguous boundaries， we proposed a novel glial tumor segmentation network，MSA-FPN，which integrates multimodal MRI inputs with an enhanced Feature Pyramid Network. The proposed architecture incorporates Receptive Field Modules to capture global context by combining diverse convolutional and dilated convolutional operations across multiple scales. Furthermore， a Multi-scale Feature Aggregation Module and Convolutional Block Attention Module were introduced to reduce feature redundancy across scales while promoting complementary information fusion. Extensive evaluations on the BraTS202l and MSD Brain public datasets demonstrate the superiority of our method. Experimental results reveal that MSA-FPN achieves significant improvements in segmenting ambiguous tumor boundaries compared to existing approaches. Specifically，on BraTS2O21，the mean Dice similarity coefficient for whole tumor and enhancing tumor regions reaches 90.69% and 86.68% ，respectively， demonstrating 0.22%～14.17% and 0.77%～12.97% enhancements over state-of-the-art methods. Consistent performance was further validated on MSD Brain， underscoring the robustness and generalizability of our framework in clinical scenarios.

Keywordsmultimodal magnetic resonance imaging； glioma segmentation； scale feature difference; attention mechanism; deep learning

0引言

胶质瘤是最常见的原发性颅脑肿瘤，约 80% 的恶性脑肿瘤为神经胶质瘤，其发病率高、死亡率高、治愈率低，且极易复发[]。（剩余13982字）