Could normalization improve robustness of abdominal MRI radiomic features?

Area of Science:

• Medical Imaging
• Radiomics
• Oncology

Background:

• Radiomics enhances cancer patient management, but generalizability is limited by multi-center, multi-scanner data.
• Image and feature normalization are proposed to address reproducibility issues in radiomics.

Purpose of the Study:

• To evaluate the impact of various image and feature normalization methods on radiomics feature robustness.
• To assess the performance of normalization techniques using a multicenter, multi-scanner abdominal MRI dataset.

Main Methods:

• Retrospective collection of 88 rectal MRIs from 3 institutions (4 scanners).
• Analysis of 93 radiomics features from six 3D regions of interest on the obturator muscle.
• Application and comparison of min-max, 1st-99th percentiles, 3-Sigma, z-score, mean centering, histogram normalization, Nyul-Udupa, and ComBat harmonization.

Main Results:

• Image normalization methods generally reduced intensity variability but often worsened feature robustness.
• Z-score image normalization slightly improved feature similarity from 9/93 to 10/93.
• Feature normalization methods, particularly 3-sigma, z-score, and ComBat, substantially increased feature similarity to 79/93.
• No image normalization method strongly increased the number of statistically similar features.

Conclusions:

• Feature normalization, especially ComBat, z-score, and 3-sigma, is crucial for enhancing radiomics feature robustness in multicenter studies.
• Image normalization alone is insufficient to guarantee the generalizability of radiomics findings across different imaging environments.