Harmonizing the consistency of liver radiomic features across multiple CT devices
Objective To evaluate the consistency of radiomic features of the healthy liver across multiple CT scanners and to investigate the impact of image-based and feature-based harmonization methods on the results.Methods Abdominal CT examinations of 243 healthy adults(Optima CT660:n=83,Revolution 512 CT:n=56,Emotion 16:n=69,Definition AS+CT:n=35)were retrospectively collected from January 1,2015 to January 1,2023 at four CT scanners of Ganzhou People's Hospital.For each patient,a 30 mm diameter three-dimensional region of interest was delineated in the liver parenchyma at the portal vein level,and 93 radiomic features were extracted using Pyradiomics.Mean centering,Z-Score,resampling,histogram matching,and ComBat methods were used to harmonize inter-device differences.The Mann-Whitney U test was used to compare the consistency of features between two different scanners before and after applying harmonization methods,and Cohen's d value was calculated to assess the effect size of different methods.Results The overall consistency of liver radiomic features was 55.38%,with 87.10%consistency among devices from the same manufacturer and 39.52%among devices from different manufacturers.After harmonization with mean centering,Z-Score,resampling,histogram matching,and ComBat methods,the proportions of consistent features were 44.82%,68.82%,66.49%,76.52%,100%,respectively;the d values were-0.57,0.62,0.57,0.78,1.59,respectively.Conclusion The consistency of liver CT radiomic features is poor between different devices,maintaining good consistency only among scanners from the same manufacturer.Image-based and feature-based harmonization methods can effectively reduce feature variation caused by different CT manufacturers and device models.Among these methods,resampling and histogram matching depend on specific parameter settings and the selection of reference images.The feature-based ComBat harmonization performs best,ensuring that all features remain consistent across different devices,which positively impacts future cross-device or cross-center studies.
万方数据
维普
