A novel gamma GLM approach to MRI relaxometry comparisons

Area of Science:

• Medical Imaging
• Biostatistics
• Quantitative MRI

Background:

• MRI relaxometry (T1, T2, R1, R2) data often exhibits constant coefficient of variation (CV) but nonconstant absolute variance.
• Standard linear models, such as ordinary least squares (OLS), assume constant variance, which can lead to erroneous conclusions when applied to MRI relaxometry data.

Purpose of the Study:

• To demonstrate that standard linear models (OLS) produce erroneous conclusions with MRI relaxometry data due to inherent CV.
• To introduce and validate a gamma generalized linear model identity link (GGLM-ID) framework that incorporates CV into parameter estimates.
• To evaluate the performance of GGLM-ID compared to OLS for contrast agent relaxivity, R1 repeatability, and liver iron content (LIC) analysis.

Main Methods:

• Eight statistical models, including OLS and GGLM-ID, were fitted to data from sulfated dextran iron oxide (SDIO) nanoparticles.
• Simulations (resampling and Monte Carlo with/without concentration error) were conducted to assess mean square error (MSE) and type I error rates.
• Model performance was evaluated on R1 repeatability and LIC datasets.

Main Results:

• OLS exhibited 4-5x higher MSE and a 20-30% type I error rate compared to GGLM-ID, which remained near the nominal 5% level in relaxivity studies.
• OLS incorrectly identified significant MRI facility effects on relaxivity, while GGLM-ID provided more consistent results.
• GGLM-ID demonstrated superior performance over OLS in modeling liver iron content (LIC).

Conclusions:

• Ordinary least squares (OLS) analysis of MRI relaxometry data can lead to erroneous conclusions.
• The proposed gamma generalized linear model identity link (GGLM-ID) framework effectively accounts for the inherent coefficient of variation (CV) in MRI data.
• GGLM-ID provides more reproducible and reliable conclusions for MRI relaxometry analyses, including relaxivity, repeatability, and liver iron content quantification.