基于轴-Transformer的医学图像分割模型Axial-TransUNetOA
Axial-TransUNet of Medical Image Segmentation Model Based on Axis-Transformer
针对TransUNet中Transformer自注意力机制计算复杂度高、捕获位置信息能力不足的问题,提出一种基于轴向注意力机制的医学图像分割网络Axial-TransUNet.该网络在保留TransUNet网络编码器、解码器以及跳跃连接的基础上,使用基于轴向注意力机制的残差轴向注意力块代替TransUNet的Transformer层.实验结果表明,在多个医学数据集上,相较于TransUNet等其他医学图像分割网络,Axial-TransUNet的Dice系数、交并比IoU有更好的表现.与TransUNet相比,Axial-TransUNet网络的参数量与浮点运算数(FLOPs)分别降低 14.9%和 30.5%.可见,Axial-TransUNet有效降低了模型复杂度,并增强了模型捕获位置信息的能力.
A medical image segmentation network Axial-TransUNet based on Axial Attention Mechanism is proposed to address the issues of high computational complexity and insufficient ability to capture positional information in the Transformer Self-Attention Mechanism in TransUNet.On the basis of retaining the TransUNet network encoder,decoder,and skip connections,this network uses residual axial attention blocks based on Axial Attention Mechanism to replace the Transformer layer of TransUNet.The experimental results show that compared to other medical image segmentation networks such as TransUNet,Axial TransUNet performs better in Dice coefficient and intersection union ratio on multiple medical datasets.Compared with TransUNet,the parameter count and FLOPs of the Axial TransUNet network are reduced by 14.9%and 30.5%,respectively.It can be seen that Axial TransUNet effectively reduces model complexity and enhances the model's ability to capture positional information.
刘文科;刘琳;韩子逸;张媛媛
青岛理工大学 信息与控制工程学院,山东 青岛 266520
计算机与自动化
医学图像分割卷积神经网络位置信息计算复杂度轴向注意力机制
medical image segmentationConvolutional Neural Networkspositional informationcomputational complexityAxial Attention Mechanism
《现代信息科技》 2024 (016)
28-33 / 6
大学生创新训练项目(202310429355);国家自然科学基金项目(61902430)
评论