基于扩散常微分方程的医学图像异常检测  

Medical image anomaly detection based on diffusion ordinary differential equations

作　　者：胡显耀 靳聪明[1] HU Xianyao;JIN Congming(School of Science,Zhejiang Sci-Tech University,Hangzhou 310018,China)

机构地区：[1]浙江理工大学理学院,杭州310018

出　　处：《浙江理工大学学报（自然科学版）》2024年第6期851-860,共10页Journal of Zhejiang Sci-Tech University(Natural Sciences)

基　　金：国家自然科学基金项目(11571314)。

摘　　要：为了识别与正常生理状态显著偏离的异常医学图像,基于异常图像相关特征往往分布于特征空间低密度区域的假设,提出了一种神经网络辅助的基于扩散常微分方程(Ordinary differential equations,ODE)的医学图像异常检测方法。该方法首先利用扩散ODE估计图像特征的似然值;然后构建神经网络,拟合图像特征在不同时刻由扩散ODE估计的似然值;最后通过扩散ODE估计的似然值和神经网络估计的似然值的加权平均得到异常分数,异常分数较大的图像被认定为异常图像。此外,为了确定异常图像的异常区域,提出了一种基于图像重构的异常定位方法,通过计算重构误差来定位异常区域。在BraTS2021脑肿瘤数据集和X射线胸透数据集上进行数值实验,结果表明该异常检测方法的异常检测性能大幅优于现有方法,且具有较好的鲁棒性。该研究提出的无监督医学图像异常检测方法和异常区域定位方法可为临床诊疗提供丰富的信息支持,有望减轻医生的工作量。To identify abnormal medical images that deviate significantly from normal physiological states,a medical image anomaly detection method,which is based on diffusion ordinary differential equations(ODE)assisted by neural network,is proposed according to the assumption that the relevant features of abnormal images often appear in low-density regions of the feature distribution.Firstly,diffusion ODE is utilized to estimate the likelihood value of the image features;then,a neural network is constructed to fit the likelihood values of the image features at different times estimated by the diffusion ODE;finally,the anomaly score of this method is the weighted average of the likelihood values estimated by the diffusion ODE and the likelihood values estimated by the neural network.Images with high anomaly scores are identified as abnormal images.In addition,an anomaly localization method based on image reconstruction is proposed to determine the abnormal regions of the abnormal images,and the reconstruction errors are calculated to locate the abnormal regions.The numerical experimental results on the BraTS2021 brain tumor dataset and the chest X-ray dataset show that the anomaly detection performance of this method is significantly better than that of other existing methods and has preferable robustness.The approach for unsupervised medical image anomaly detection and the method to locate the abnormal regions proposed in this article can provide mass information support for clinical diagnosis and treatment,and are expected to reduce the workload of doctors.

关 键 词：异常检测 扩散常微分方程 神经网络 医学图像 异常定位 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]