Predicting treatment efficacy via quantitative magnetic resonance imaging: a Bayesian joint model

Area of Science:

• Neuro-oncology
• Medical Imaging
• Biostatistics

Background:

• High-grade gliomas have a poor prognosis with a median survival of only one year.
• Current methods for assessing treatment efficacy are delayed, occurring 5-6 months post-diagnosis.

Purpose of the Study:

• To develop a predictive model for high-grade glioma treatment efficacy using quantitative magnetic resonance imaging (qMRI) data.
• To evaluate the performance of this predictive model in assessing 1-year survival status.

Main Methods:

• A joint, two-stage Bayesian model was developed.
• Stage I involved smoothing image data using a multivariate spatiotemporal pairwise difference prior.
• Stage II utilized four summary statistics from Stage I within a generalized non-linear model (probit link, MARS basis) for survival prediction, employing Gibbs sampling and RJMCMC.

Main Results:

• The proposed model achieved higher overall correct classification rates compared to other methods.
• Accounting for spatiotemporal correlation in MRI data improved prediction accuracy.
• The generalized non-linear model provided a flexible decision boundary, enhancing predictive performance.

Conclusions:

• Quantitative MRI can predict high-grade glioma treatment efficacy as early as 3 weeks post-therapy initiation.
• This early assessment enables timely initiation of salvage treatments, potentially improving patient survival.
• The developed Bayesian model effectively leverages qMRI data for accurate survival prediction in high-grade glioma patients.