Intensity harmonization techniques influence radiomics features and radiomics-based predictions in sarcoma patients

Area of Science:

• Medical Imaging
• Radiomics
• Machine Learning

Background:

• Multicentric MRI data require intensity harmonization for quantitative analysis due to non-standardized signal intensities.
• Radiomics, combining radiological phenotype quantification with machine learning, aims to improve predictive models for patient outcomes.

Purpose of the Study:

• To evaluate the impact of five intensity harmonization techniques (IHTs) on radiomics features and their correlation with metastatic relapse-free survival (MFS) in sarcoma patients.
• To assess the performance variation of radiomics-based models for predicting 2-year metastatic relapse based on different IHTs.

Main Methods:

• Post-processing of 70 sarcoma patients' T2-weighted MRIs using five IHTs (IHTstd, IHTfat, IHTHM.1, IHTHM.All, IHTHM.All.C) and a no-IHT dataset.
• Extraction of 45 radiomics features (RFs) for each dataset.
• Unsupervised clustering and multivariate Cox models to correlate clusters with MFS.
• Development of radiomics-based supervised models to predict 2-year metastatic relapse using a training set of 50 patients.

Main Results:

• Intensity harmonization significantly influenced all extracted radiomics features (p < 0.0001-0.02).
• Only clusters from No-IHT, IHTstd, IHTHM.All, and IHTHM.All.C datasets showed significant correlation with MFS (p = 0.004-0.02).
• Radiomics model performance for predicting 2-year metastatic relapse varied significantly across IHTs, with AUROC ranging from 0.688 (IHTstd) to 0.823 (IHTHM.1).

Conclusions:

• Intensity harmonization techniques profoundly affect radiomics feature values and the reproducibility of predictive models in sarcoma.
• The choice of IHT is critical and must be carefully detailed in radiomics post-processing pipelines to ensure reliable and reproducible analyses.
• Careful selection and reporting of intensity harmonization methods are essential for robust radiomics research and clinical application.