摘要：

磁共振成像(MRI)脑组织分割在临床诊断、治疗规划、神经学研究、脑功能定位等方面具有重要意义，有助于帮助医生理解和治疗各种神经系统相关的疾病。目前基于Transformer的方法基于自注意力机制提取特征并进行分割，存在为降低自注意力机制的复杂度而牺牲分割精度的问题。为了解决上述问题，提出基于查询自适应双层自注意力机制的MRI脑组织分割网络。在编码器-解码器的网络架构中将查询自适应双层自注意力模块作为主编码器。查询自适应双层自注意力模块包括动态稀疏的粗粒度层和像素级自注意力的细粒度层。粗粒度层基于自注意力计算描述图像块间的相关性，动态地过滤掉不相关的图像块，实现灵活高效的计算，细粒度在挑选出的相关区域应用像素到像素的自注意力，以提升分割精度。该模块可提升分割精度，同时有效控制计算复杂度。分割算法在常用的脑核磁共振图像分割基准上进行验证，实验结果表明，该方法的Dice系数(DSC) 为0.917±0.030，豪斯多夫距离(HD95)为1.196±0.613 mm，在图像分割性能上优于其他先进方法，证明了该方法在核磁共振图像脑组织分割任务中的有效性和准确性。

关键词: 磁共振成像, 脑组织分割, 深度学习, 自注意力, 分割精度

Abstract:

Brain tissue segmentation in Magnetic Resonance Imaging (MRI) is vital in neuroimaging analysis applications, such as diagnosis, planning, and research. Current transformer methods use the self-attention mechanism to extract features for segmentation, and their accuracy needs to be improved when the computational complexity is high. To address this issue, this paper proposes a network for MRI brain tissue segmentation that combines a query-adaptive bilevel self-attention module in a U-shaped architecture. The query-adaptive bilevel self-attention module consists of a sparse coarse-grained layer and a pixel-level self-attention layer to balance accuracy and complexity. Specifically, the coarse-grained layer utilizes self-attention to dynamically filter out irrelevant image blocks for more flexible and efficient computation, whereas the fine-grained layer applies pixel-to-pixel self-attention for high-precision segmentation. The module achieves high performance while restricting the computation cost. The algorithm is validated on popular brain MRI segmentation benchmarks and outperforms state-of-the-art methods with a Dice Similarity Coefficient (DSC) of 0.917±0.030 and HD95 of 1.196±0.613 mm. Experimental results demonstrate the effectiveness and accuracy of the algorithm in segmenting brain tissue for MRI.

Key words: Magnetic Resonance Imaging(MRI), brain tissue segmentation, deep learning, self-attention, segmentation accuracy