融合残差结构和注意力机制的下颌骨CT图像分割方法  

Mandible CT image segmentation method fused with residual structure and attention mechanism

作　　者：方梓涛 刘丹[1] 吴扬东[1] 何玲[1] FANG Zitao;LIU Dan;WU Yangdong;HE Ling(Key Laboratory of Advanced Manufacturing Technology,Ministry of Education,Guizhou University,Guiyang 550025,China)

机构地区：[1]贵州大学现代制造技术教育部重点实验室,贵阳550025

出　　处：《智能计算机与应用》2024年第2期83-89,共7页Intelligent Computer and Applications

摘　　要：针对传统深度学习方法在下颌骨CT图像分割中存在的问题,本文提出一种融合残差结构和注意力机制的改进Unet网络。将注意力机制融入Unet的解码器,构建上采样注意力模块,在不同维度上进行平均池化和最大池化,使网络更加关注下颌骨区域的信息,增强下颌骨分割能力;将残差结构融入Unet网络的编码器,解决深度网络训练时的网络退化和梯度消失问题;采用迁移学习训练的方法,避免因下颌骨图像数据不足导致的网络收敛慢的问题。对比实验表明,改进Unet网络平均交并比达到94.68%,各评价指标均优于FCN、DeeplabV1和SegNet网络。Aiming at the problems existing in traditional mandible CT image segmentation methods,this paper proposes an improved Unet network that combines residual structure and attention mechanism.By integrating the attention mechanism into Unet decoder,the designed model constructs an upsampling attention module and performs average pooling and maximum pooling in different dimensions,making the network pay more attention to the information of the mandible region and enhance the ability of mandible segmentation.The residual structure is introduced into the encoder to solve the problem of network degradation and gradient disappearance during deep network training.The transfer learning is used to avoid the problem of slow convergence caused by insufficient mandible image data.Comparative experiments show that the Mean Intersection over Union of the improved Unet network reaches 94.68%,and all evaluation indicators are better than FCN,DeeplabV1 and SegNet networks.

关 键 词：下颌骨 Unet 残差结构 注意力机制 迁移学习 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]