Standardization Methods For Multi-center Medical Imaging Data

In recent years,medical big data analysis based on artificial intelligence has received quite a lot of attention.Specifically,radiomics has become a hot research topic.With the increase in development and improvement of radiomics,it has become crucial to have sufficient data to ensure that radiomics analysis is more general and universal.The data obtained by a single medical institution is often small,driving people to integrate data from multiple medical institutions to form a multi-center data set.However,because multi-center data uses different collection devices and sets different collection parameters during the collection process,there are huge differences in data storage formats,specifications,brightness,and contrast.These differences will interfere with a follow-up clinician's differential diagnosis and radiomics analysis.To this end,this article proposes a set of multi-center data standardization schemes using an MRI-based glioma image data as entry point,in-depth analysis of key issues such as information enhancement and data unification in multi-center data standardization and at the same time puts forward a certain optimization scheme to reduce the difference in data distribution,improve adaptability and robustness of data,and mitigate impact of multi-center data.The main research contents of this article are as follows:(1)Aiming at the problem of huge difference in contrast and brightness of different images in the current multi-center image data standardization process,this thesis proposes an automatic histogram specification method in combination with grid search(HSASR).HSASR can uniformly regulate the dynamic range of contrast and brightness of an overall data under the action of an appropriate reference frame and at the same time effectively enhance tumor area information.In the previous regularization process,radiologists have generally selected reference frames for mapping;but as the amount of data increased,the disadvantages of manual selection such as heavy workload and extreme time consumption were gradually exposed.There was also no guarantee of best choice for a reference frame.HSASR canautomatically select the optimal reference image template to standardize image contrast and brightness distribution.Finally,the data that is processed by this method is verified by radiomics.Experiment results show that the data processed by this method has an average increase of 13% in performance indexes such as AUC and accuracy in the prediction of a classification effect.(2)Furthermore,in view of the problem that an image that is processed by a data normalization method is blurring domain boundaries,a method based on multi-peak histogram(MPH)is proposed.Although HSASR unifies the overall image brightness range and highlights the contrast between a tumor and other tissues,blurring of the boundaries of divided areas of the tumor in some image data is more prominent;for example,for gliomas,the tumor area can be divided generally into enhancement,necrosis,edema,etc.Different areas jointly determine the grade of a glioma.In order to provide more abundant tumor information for a study,MPH emphasizes on the information from each tumor area on the basis of ensuring brightness consistency of a multi-center data image.Experiment results show that the evaluation index of MPH increases by 2% on HSASR,thus verifying the effectiveness of this method.