基于高斯偏置自注意力和交叉注意力的医学图像分割模型

Gaussian-bias Self-attention and Cross-attention Based Module for Medical Image Segmentation

为解决医学图像分割中目标之间存在特征差异、不同切片图像中存在同一解剖结构的相似表征和器官与背景的区分度低造成冗余信息过多的问题,提出了一种基于高斯偏置自注意力和交叉注意力的网络模型(Gaussian bias and Contextual cross Attention U-Net,GCA-UNet)。采用残差模块建立空间先验假设,通过高斯偏置自注意力&外注意力模块的高斯偏置自注意力来学习空间先验假设和强化相邻区域的特征表示,并利用外注意力机制学习同一样本下不同切片之间的相关性;上下文交叉注意力门控利用多尺度特征提取来强化结构和边界信息,同时对上下文语义信息进行重新校准并筛除冗余信息。实验结果表明,在Synapse腹腔CT多器官分割数据集和ACDC心脏MRI数据集上,GCA-UNet网络的分割精度指标Mean Dice分别达到了81.37%和91.69%,在Synapse数据集上边界分割精度指标Mean hd95达到16.01。相比其他先进医学影像分割模型,GCA-Unet分割精度更高,具有更清晰的组织边界。

To address the problems in medical image segmentation,such as varying target sizes,diverse representations of the same anatomical structures across slices,and low distinction between organs and background leading to excessive redundant information,a novel model based on Gaussian bias self-attention and contextual cross attention,named Gaussian bias and cross attention U-Net(GCA-UNet),is proposed.The model utilizes residual modules to establish spatial prior hypotheses,employs Gaussian bias self-attention and external attention mechanisms to learn spatial priors and enhance feature representations of adjacent areas,and uses external attention to understand inter-sample correlations.The cross attention gated mechanism leverages multi-scale feature extraction to reinforce structural and boundary information while recalibrating contextual semantic information and filtering out redundant data.Experimental results on the Synapse abdominal CT multi-organ segmentation dataset and ACDC cardiac MRI dataset show that,the proposed GCA-UNet achieves Mean Dice accuracy metrics of 81.37% and 91.69%,respectively,with a Mean hd95 boundary precision of 16.01 on the Synapse dataset.Compared to other advanced medical image segmentation models,GCA-UNet offers higher segmentation accuracy with clearer tissue boundaries.