Influence of Magnetic Field Strength on Magnetic Resonance Imaging Radiomics Features in Brain Imaging, an In Vitro and In Vivo Study

Samy Ammari, Stéphanie Pitre-Champagnat, Laurent Dercle, Émilie Chouzenoux, Salma Moalla, Sylvain Reuzé, Hugues Talbot, T. Mokoyoko, Joya Hadchiti, Sébastien Diffetocq, Andreas Volk, Mickeal El Haik, Sara Lakiss, Corinne Balleyguier, Nathalie Lassau, François BidaultFrontiers in Oncology 2021 · DOI

TL;DR

This study reveals that magnetic field strength (1.5T vs. 3T) significantly impacts radiomic features extracted from MRI scans, meaning data acquired at different field strengths cannot be directly compared. This highlights a critical need for standardization in MRI radiomics to ensure reproducible and reliable clinical applications.

What Problem Does This Paper Address?

Radiomics, the extraction of quantitative features from medical images, holds immense promise for personalized medicine, particularly in oncology. It aims to uncover imaging biomarkers invisible to the naked eye that can predict tumor characteristics, treatment response, and patient outcomes. However, a major hurdle for radiomics' widespread clinical adoption is the lack of standardization. Unlike X-ray CT, MRI images are highly dependent on scanner manufacturer, sequence type, and acquisition parameters. This variability can lead to inconsistent radiomic feature values, compromising reproducibility. This paper specifically addresses the question of how magnetic field strength, a fundamental parameter in MRI, influences these radiomic features, particularly texture analysis, which is crucial for characterizing tissue heterogeneity.

Methods

The researchers employed a multi-faceted approach using both in vitro phantoms and in vivo human data.

  • Phantoms: They created two types of phantoms:
  • Homogeneous phantoms: Mimicking cerebrospinal fluid and blood vessels, these phantoms contained varying concentrations of gadolinium.
  • Heterogeneous phantoms: Designed to represent brain white matter, these phantoms incorporated agarose gel with polystyrene beads of different sizes.
  • In Vivo Data: Ten healthy volunteers underwent MRI scans.
  • MRI Acquisition: T1-weighted 3D SPGR sequences were acquired on both 1.5T and 3T MRI devices from the same manufacturer. Crucially, the same acquisition parameters (matrix size, field of view, slice thickness) were used for both field strengths, with the exception of pixel size, which was varied for phantom studies.
  • Texture Analysis: The LIFEX software was used to segment regions of interest (ROIs) and extract 38 radiomic features, including first-order and second-order statistics derived from histogram analysis and texture matrices (co-occurrence matrix, gray-level run length matrix, gray-level zone length matrix).
  • Statistical Analysis: Student's t-tests were used to compare feature values between 1.5T and 3T. Spearman's rho correlation coefficients were employed to assess the association between acquisition parameters and textural changes.

Datasets

  • In Vitro Datasets:
    • Homogeneous phantoms with varying gadolinium concentrations (0.25 to 2 mmol/l).
  • Heterogeneous phantoms with polystyrene beads of different sizes (1, 2, and 3 mm) embedded in agarose gel.
  • In Vivo Dataset: MRI scans from 10 healthy volunteers.

The study did not introduce new benchmark datasets but rather used controlled phantom setups to isolate the effect of magnetic field strength and other parameters.

Key Insights and Ideas

  • Field Strength is a Major Confounder: The most significant takeaway is that radiomic features, especially texture features, are highly sensitive to magnetic field strength. Data from 1.5T and 3T scanners are not directly interchangeable.
  • Texture is Particularly Affected: Texture analysis, which quantifies image heterogeneity, appears to be more susceptible to field strength variations than simpler first-order statistics.
  • Beyond Field Strength: While field strength is a primary driver of differences, the study also confirmed that gray-level resampling and pixel size (spatial resolution) can independently influence radiomic feature values.
  • Phantom Utility: The use of well-characterized phantoms allowed for the isolation and systematic evaluation of specific parameters like gadolinium concentration and bead size, providing a controlled environment to understand their impact on radiomics.

Key Results

  • Phantom Study: In the phantom dataset, 10 out of 15 radiomic features were significantly different between 1.5T and 3T scans (p < 0.05).
  • In Vivo Study: These findings were validated in healthy volunteers, where 15 out of 38 texture features showed significant differences between 1.5T and 3T scans for the same anatomical regions.
  • Resampling Impact: Gray-level resampling significantly impacted texture parameters extracted from matrices, with correlation decreasing as the number of gray levels decreased.
  • Pixel Size Impact: Pixel size (influenced by matrix size and FOV) altered radiomic output in various ways, affecting different features on either 1.5T, 3T, or both.

Limitations and Future Work

  • Single Manufacturer: The study used MRI devices from only one manufacturer. Future work should investigate if these findings hold true across different manufacturers.
  • Limited Volunteer Pool: While 10 volunteers were included, a larger and more diverse cohort would strengthen the generalizability of the in vivo findings.
  • Standardization Focus: The authors emphasize the need for standardization strategies for radiomics in MRI. Future work should focus on developing and validating methods to harmonize radiomic features across different field strengths and acquisition protocols. This could involve developing correction algorithms or establishing specific protocols for each field strength.
  • Clinical Translation: While the study provides crucial insights for clinical practice, further research is needed to directly link these findings to clinical outcomes and diagnostic accuracy.

Summary updated

Abstract

According to daily used protocols for clinical examinations, radiomic features extracted on 1.5T should not be used interchangeably with 3T when evaluating texture features. Such confounding factor should be adjusted when adapting the results of a study to a different platform, or when designing a multicentric trial.
BibTeX
@inproceedings{ammari2021influence,
  author = {Samy Ammari and Stéphanie Pitre-Champagnat and Laurent Dercle and Émilie Chouzenoux and Salma Moalla and Sylvain Reuzé and Hugues Talbot and T. Mokoyoko and Joya Hadchiti and Sébastien Diffetocq and Andreas Volk and Mickeal El Haik and Sara Lakiss and Corinne Balleyguier and Nathalie Lassau and François Bidault},
  title = {{Influence of Magnetic Field Strength on Magnetic Resonance Imaging Radiomics Features in Brain Imaging, an In Vitro and In Vivo Study}},
  booktitle = {Frontiers in Oncology},
  year = {2021},
  doi = {10.3389/fonc.2020.541663},
}
Figure