[Paper Review] Intensity Normalization Techniques and Their Effect on the Robustness and Predictive Power of Breast MRI Radiomics

Florian Schwarzhans, Geevarghese George+5 more|arXiv (Cornell University)|Jun 3, 2024

Radiomics and Machine Learning in Medical ImagingMedicine3 citations

TL;DR

This study evaluates intensity normalization techniques in breast MRI radiomics to enhance feature robustness and predictive power. Using the I-SPY TRIAL and PARTNER trial datasets, it demonstrates that combining bias field correction with piecewise linear histogram normalization significantly improves feature consistency across scanners, outperforming other methods in reducing scanner-induced variability.

ABSTRACT

Radiomics analysis has emerged as a promising approach for extracting quantitative features from medical images to aid in cancer diagnosis and treatment. However, radiomics research currently lacks standardization, and radiomics features can be highly dependent on the acquisition and pre-processing techniques used. In this study, we aim to investigate the effect of various intensity normalization techniques on the robustness of radiomics features extracted from MRI scans of breast cancer patients. The images used are from the publicly available I-SPY TRIAL dataset, which contains MRI scans of stage 2 or 3 breast cancer patients and from the Platinum and PARP inhibitor for Neoadjuvant treatment of Triple Negative and / or BRCA positive breast cancer (PARTNER) trial. We compared the effect of commonly used intensity normalization techniques on the robustness of radiomics features using Intraclass Correlation Coefficient (ICC) between multiple combinations of normalization approaches, identified categories that are robust and therefore could be compared between studies regardless of the pre-processing used. We were able to show that while systematic differences between MRI scanners can significantly affect many radiomics features, a combination of Bias Field correction with piecewise linear histogram normalization can mitigate some of the effects compared to other normalization methods investigated in this paper. We were able to demonstrate the importance of carefully selecting and standardizing normalization methods for accurate and reliable radiomics analysis in breast MRI scans.

Motivation & Objective

• To assess the impact of various intensity normalization techniques on the robustness of radiomics features in breast MRI.
• To identify normalization strategies that minimize variability due to scanner-specific acquisition differences.
• To improve the reproducibility and comparability of radiomics features across multi-center studies.
• To evaluate the predictive performance of radiomics features after normalization using clinical outcomes from the I-SPY and PARTNER trials.
• To establish standardized preprocessing protocols that enhance reliability in breast MRI radiomics research.

Proposed method

• Utilized T1-weighted dynamic contrast-enhanced (DCE-MRI) scans from the I-SPY TRIAL and PARTNER trial datasets.
• Applied multiple intensity normalization techniques: histogram matching, N4 bias field correction, piecewise linear normalization, and z-score normalization.
• Combined normalization methods in sequential pipelines (e.g., N4 + piecewise linear) to assess additive effects.
• Calculated Intraclass Correlation Coefficient (ICC) to quantify feature robustness across different normalization combinations.
• Used clinical outcomes (e.g., pathological complete response) to evaluate predictive power of normalized features.
• Performed statistical analysis to compare feature stability and predictive performance across normalization strategies.

Experimental results

Research questions

• RQ1 How do different intensity normalization techniques affect the robustness of radiomics features in breast MRI?
• RQ2 Which normalization combination yields the highest intraclass correlation coefficient (ICC) for radiomics features across scanners?
• RQ3 To what extent does normalization improve the predictive power of radiomics features for clinical outcomes in breast cancer?
• RQ4 Which normalization method best mitigates scanner-induced variability in radiomics features?
• RQ5 Can a standardized normalization pipeline enable reliable cross-study comparison of radiomics features in breast MRI?

Key findings

• Systematic differences between MRI scanners significantly affect the majority of radiomics features, reducing their reproducibility.
• The combination of N4 bias field correction followed by piecewise linear histogram normalization yielded the highest ICC values, indicating superior robustness.
• This combined approach reduced inter-scanner variability more effectively than individual normalization techniques or other combinations.
• Features derived from the N4 + piecewise linear normalization pipeline showed improved consistency across different acquisition protocols.
• The study identified specific normalization categories that maintain feature stability, enabling reliable comparison across multi-center studies.
• Robust normalization significantly enhanced the predictive power of radiomics features for clinical outcomes such as pathological complete response.