计算机技术与发展2025，Vol.35Issue(10)：53-61,9.DOI:10.20165/j.cnki.ISSN1673-629X.2025.0130
多尺度融合与位置增强的多模态脑肿瘤分割模型
Multimodal Brain Tumor Segmentation Model Based on Multi-scale Fusion and Position Enhancement
摘要
Abstract
Multimodal brain tumor images have comprehensive information,which can provide doctors with more comprehensive and accurate information.However,due to the lack of brain tumor image data,the complex and varied brain structure,the difference of tumor morphology and the close interweaving state with normal tissues,these comprehensive factors together constitute the severe challenges faced by brain tumor segmentation models in effectively utilizing and deeply mining the key information of multi-modal images.Therefore,we propose a multimodal brain tumor segmentation model based on multi-scale fusion and position enhancement.Firstly,a residual pre-activation module is designed to replace the coding and decoding layers in U-Net to preserve the details of low-level features.Secondly,a multi-scale cavity fusion attention module is designed in the bottleneck layer to conduct multi-source sampling of global and local information of brain tumors,aiming to effectively utilize the comprehensive information between multi-modal medical images.Finally,the coordinate information enhancement module is embedded in the jump connection layer to fully mine the spatial location information of the brain tumor region.Experimental results on the BraTS2018 and BraTS2019 datasets show that the Dice coefficients of the proposed network for WT,TC and ET segmentation are 91.64％,90.28％and 88.57％,respectively,and the Sensitivity coefficient and Hausdorff distance are also optimal compared with other networks.It is showed that the designed network can effectively utilize lesion information presented by different modal images,and can fuse tumor features of different scales,effectively improving segmentation accuracy.关键词
医学图像分割/脑部肿瘤/多尺度/多模态信息/注意力机制Key words
medical image segmentation/brain tumor/multi-scale/multimodal information/attention mechanism分类
信息技术与安全科学引用本文复制引用
史雯雯,刘石坚,邹峥..多尺度融合与位置增强的多模态脑肿瘤分割模型[J].计算机技术与发展,2025,35(10):53-61,9.基金项目
国家自然科学基金(62172095) （62172095）
福建省科技厅自然科学基金项目(2022J01932) （2022J01932）
福建省教育厅科研项目(JAT210283) （JAT210283）
