基于空间自注意力机制和深度特征重建的脑MR图像分割方法

Brain MR Image Segmentation Based on Spatial Self-attention Mechanism and Depth Feature Reconstruction

准确分割核磁共振(magnetic resonance,MR)图像中的脑组织是临床诊断、手术计划和辅助治疗的关键步骤.深度学习在各种图像分割任务中表现出巨大潜力,现有模型没有一种有效方法汇总远距离像素间的关系.在网络解码阶段不能很好地融合不同层级的特征,导致无法准确定位.为克服上述问题,本文提出一种基于空间自注意力机制和深度特征重建的脑MR图像分割方法,构建了一个可以融合3维信息的2D模型,可快速准确对3D结构图像进行密集预测.在MRBrainS13数据集和IBSR数据集上进行充分地实验研究,结果表明本文方法在3D多模态和单模态脑MR图像分割方面优于目前的2D模型,运算和推理时间相比3D模型小很多,性能却十分接近.

Accurate segmentation of brain tissue in MR images is a key step in clinical diagnosis,surgical planning and adjuvant treatment.Deep-learning shows great potential in various image segmentation tasks,and existing models do not have an effective way to summarize the relationship between long-distance pixels.In the network decoding stage,the features of different levels cannot be well integrated,resulting in the inability to accurately locate.To overcome the above problems,this paper proposes a brain MR image segmentation method based on spatial self-attention mechanism and depth feature reconstruction,and constructs a 2D model that can fuse 3D information,which can quickly and accurately perform dense prediction on 3D structural images.The proposed method is fully experimented on MRBrainS13 data sets and IBSR data sets,and the results show that the model outperforms the current 2D model in 3D multimodal and unimodal brain MR image segmentation,with less computing and inference time compared to the 3D model,whereas the performance is very close.