摘要
在核医学中,单光子发射计算机断层(single-photon emission computed tomograpy,SPECT)骨显像是辅助医师诊断癌症的重要手段。针对骨显像图像信噪比低、边界模糊、病灶小难以提取和人工勾画病灶耗时等问题,提出一种基于改进U-Net网络的骨显像病灶自动分割算法。该算法在U-Net的原卷积块基础上,采用了多尺度密集连接(multi-scale dense connection,MDC)的方式来提高对小病灶特征的提取能力,同时解决了网络加深后出现的梯度消失问题。其次,为提取病灶的细节特征,在密集连接和跳跃连接处引入了注意力机制结构。最后,针对使用小样本数据集,模型难以收敛的问题,采用迁移学习的方法,优化了模型的初始参数,提升模型的泛化能力和分割效率。此外,为了降低计算量、进一步提高分割效果,对数据集进行了裁剪和去噪。同时,将处理后的图像采用旋转、镜像等方法进行了数据扩充。实验结果表明,改进的U-Net的识别精确率(precision)、平均交并比(mean intersection-over-union,mIoU)分别能达到0.7352、0.4673,效果优于目前主流的分割算法,具有一定实际应用价值。
In nuclear medicine,single-photon emission computed tomography(SPECT)bone imaging is an important means to assist physicians in diagnosing diseases.Aiming at the problems of low signal-to-noise ratio,blurred boundaries,small lesions,and time-consuming manual lesion delineation in bone imaging images,an automatic segmentation algorithm for bone imaging lesions based on improved U-Net network was proposed.Based on the original convolution block of U-Net,the algorithm adopts a multi-scale dense connection(MDC)method to improve the extraction ability of small lesion features,and at the same time solves the problem of gradient disappearance after the network is deepened.Second,to extract detailed features of lesions,an attention mechanism structure is introduced at dense and skip connections.Finally,in view of the problem that the model is difficult to converge when using a small sample dataset,the transfer learning method is used to optimize the initial parameters of the model and improve the generalization ability and segmentation efficiency of the model.In addition,in order to reduce the amount of computation and further improve the segmentation effect,the dataset is cropped and denoised.At the same time,the processed images are augmented by rotation,mirroring and other methods.The experimental results show that the improved U-Net′s recognition precision and mean intersection-over-union ratio(mIoU)can reach 0.7352 and 0.4673,respectively,which are better than the current mainstream segmentation algorithms,and have certain practical application value.
作者
余泓
罗仁泽
陈春梦
罗任权
李华督
YU Hong;LUO Renze;CHEN Chunmeng;LUO Renquan;LI Huadu(College of Electrical Engineering and Information,Southwest Petroleum University,Chengdu,Sichuan 610500,China;Department of Nuclear Medicine,The No.2 People′s Hospital of Yibin,Yibin,Sichuan 644000,China)
出处
《光电子．激光》
CAS
CSCD
北大核心
2022年第10期1110-1120,共11页
Journal of Optoelectronics·Laser
基金
四川省科技计划项目(2019CXRC0027)资助项目
关键词
SPECT骨显像
多尺度密集连接(MDC)
图像处理
注意力机制
迁移学习
single-photon emission computed tomography(SPECT)bone imaging
multi-scale dense connections(MDC)
image preprocessing
attention mechanism
transfer learning
