基于融合去噪的多模态脑肿瘤 MR图像分割

摘要：

针对多模态脑肿瘤MR图像分割面临模态融合困难和融合过程中易产生噪声的问题及模态处理后全局和局部特征提取不充分造成分割精度降低的问题。首先设计了结合ViT(vision Transformer)和反卷积(transposed convolution)的模块，用于实现多模态MR图像的高效融合与深度去噪。然后创建了融合Transformer skip connection模块(TSC),利用Transformer 的多头交叉注意力机制，对传统跳跃连接进行了革新，用于捕获融合去噪后模态特征图的全局和局部特征，进一步提升分割精度。最后在公开数据集 BraTS2021上的实验结果表明，该方法在分割肿瘤时,精度达到了88.39%,Dice指数为83.44%,Hausdorff距离仅为2.3566.

The segmentation of multimodal brain tumor MR images faces the problems of difficulty in modal fusion, easy generation of noise during the fusion process, and insuficient extraction of global and local features after modal processing,resulting in reduced segmentation accuracy. This paper first designs a module(VTC) that combines vision Transformer and deconvolution to achieve efficient fusion and deep denoising of multimodal MR images. Then, a skip connection module(TSC)fused with Transformer is created. The traditional skip connection is innovated by using the muli-head cross-attention mechanism of Transformer to capture the global and local features of the modal feature map after fusion and denoising, further improving the segmentation accuracy. Finally, the experimental results on the public dataset BraTS2021 show that this method achieves an accuracy of 88. 39% when segmenting tumors, a Dice coefficient of 83. 44%. and a Hausdorff distance of only2.3566.

作者: 

申俊丽，蔺崇玉，海玉曼

Shen Junli，Lin Chongyu，Hai Yuman

机构地区：

河南师范大学计算机与信息工程学院(人工智能学院)

引用本文：

申俊丽，蔺崇玉，海玉曼。基于融合去噪的多模态脑肿瘤MR图像分割[J].河南师范大学学报(自然科学版)，2026,54(2):135-142. (Shen Junli. Lin Chongyu, Hai Yuman. Multimodal brain tumor MR image segmentation based on fusion denoising[J].Journal of Henan Normal University(Natural Science Edition),2026,54(2):135-142.

DOI:10.16366/j.cnki.1000-2367.2024.07.17.0001.)

基金：

国家自然科学基金；河南省自然科学基金；河南师范大学国家级科研项目培育基金

关键词：

脑肿瘤MR；融合去噪；U-Net；Transformer；图像分割

brain tumor MR; fusion denoising; U-Net; Transformer; image segmentation

分类号：

R318；TP391.41

基于融合去噪的多模态脑肿瘤MR图像分割.pdf