基于正逆卷积混合的肺结节分割研究  

Segmentation of Lung Nodules in CT Image Based on Combination of Convolution and Involution Operators

作　　者：史轶伦 张萍 殷雁君 于磊 徐巧枝 SHI Yi-lun;ZHANG Ping;YIN Yan-jun;YU Lei;XU Qiao-zhi(College of Computer Science and Technology,Inner Mongolia Normal University,Hohhot 010022,China;Nephrology of People's Hospital of Inner Mongolia Autonomous Region,Hohhot 010020,China)

机构地区：[1]内蒙古师范大学计算机科学技术学院,内蒙古呼和浩特010022 [2]内蒙古自治区人民医院肾脏内科,内蒙古呼和浩特010020

出　　处：《内蒙古师范大学学报（自然科学汉文版）》2023年第3期276-284,共9页Journal of Inner Mongolia Normal University(Natural Science Edition)

基　　金：内蒙古自治区自然科学基金资助项目(2021MS06031,2022ZD05);内蒙古自治区科技计划资助项目(2022YFSH0010);内蒙古自治区“草原英才”产业创新创业人才团队基金资助项目。

摘　　要：深度卷积神经网络在CT图像肺结节分割任务中取得了良好效果,但是卷积算子的空间不变性和通道特异性特点,使卷积核参数化为固定大小的矩阵,基于反向传播算法的更新在一定程度上阻碍学习到的卷积核在不同分辨率输入图像之间的传输,导致相关模型在捕获图像深层特征方面存在不足。逆卷积算子具有与卷积算子相反的特点,可弥补卷积算子的不足,将其引入CT图像肺结节分割任务,提出并设计正逆卷积混合算子,然后提出基于该算子的肺结节分割模型CIU-Net,实验证明正逆卷积混合算子的有效性。相比其他经典模型,CIU-Net具有更好的分割精度,在参数量及泛化性方面也具有优越性。Deep convolution neural networks have achieved good results in CT image lung nodule segmentation tasks.However,the spatial invariance and channel specificity of the convolution operator parameterizes the convolution kernel as a fixed size matrix,and the update based on the back-propagation algorithm prevents the learned convolution kernel from being transferred between input images of different resolutions to some extent,these lead to some shortcomings while capturing deep features of images in related models.In addition,the computational cost of convolution operators has become one of the key factors affecting the clinical application of related models.The involution operator has the opposite characteristics comparing with the convolution operator and can compensate for above deficiencies of the convolution operator.Therefore,it is introduced into the lung nodule segmentation task of CT image,and a hybrid operator consisted of convolution and involution operators is proposed and designed in the paper,and then a hybrid operator based lung nodule segmentation model CIU⁃Net is proposed.The experimental results demonstrate the effectiveness of the hybrid operator,and the CIU⁃Net has better segmentation accuracy and shows superior performance especially for small lung nodules,as well as superiority in terms of number of parameters and generalization compared with other classical models.

关 键 词：肺结节分割 逆卷积算子 卷积算子 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]