Abstract
Intensity harmonization techniques (IHT) are mandatory to homogenize multicentric MRIs before any quantitative analysis because signal intensities (SI) do not have standardized units. Radiomics combine quantification of tumors' radiological phenotype with machine-learning to improve predictive models, such as metastastic-relapse-free survival (MFS) for sarcoma patients. We post-processed the initial T2-weighted-imaging of 70 sarcoma patients by using 5 IHTs and extracting 45 radiomics features (RFs), namely: classical standardization (IHT(std)), standardization per adipose tissue SIs (IHT(fat)), histogram-matching with a patient histogram (IHT(HM.1)), with the average histogram of the population (IHT(HM.All)) and plus ComBat method (IHT(HM.All.C)), which provided 5 radiomics datasets in addition to the original radiomics dataset without IHT (No-IHT). We found that using IHTs significantly influenced all RFs values (p-values: < 0.0001-0.02). Unsupervised clustering performed on each radiomics dataset showed that only clusters from the No-IHT, IHT(std), IHT(HM.All), and IHTHM.All.C datasets significantly correlated with MFS in multivariate Cox models (p = 0.02, 0.007, 0.004 and 0.02, respectively). We built radiomics-based supervised models to predict metastatic relapse at 2-years with a training set of 50 patients. The models performances varied markedly depending on the IHT in the validation set (range of AUROC from 0.688 with IHT(std) to 0.823 with IHT(HM.1)). Hence, the use of intensity harmonization and the related technique should be carefully detailed in radiomics post-processing pipelines as it can profoundly affect the reproducibility of analyses.