Multiple Sclerosis Resource Centre

Welcome to the Multiple Sclerosis Resource Centre. This website is intended for international healthcare professionals with an interest in Multiple Sclerosis. By clicking the link below you are declaring and confirming that you are a healthcare professional

You are here

Using multiple imputation to efficiently correct cerebral MRI whole brain lesion and atrophy data in patients with multiple sclerosis

Alicia S. Chua, Svetlana Egorova, Mark C. Anderson, Mariann Polgar-Turcsanyi, Tanuja Chitnis, et al.

NeuroImage, Volume 119, 1 October 2015, Pages 81-88

Highlights
 

  • Multiple imputation (MI) was used to fill-in missing manually-corrected MRI data.
  • An imputation model with concurrent automated measures is efficient for MI.
  • Our imputation model explained a high proportion of variance in the missing MRI data.
  • MI can be used to fill-in missing data with little impact on bias or variance.

Abstract
 

Automated segmentation of brain MRI scans into tissue classes is commonly used for the assessment of multiple sclerosis (MS). However, manual correction of the resulting brain tissue label maps by an expert reader remains necessary in many cases. Since automated segmentation data awaiting manual correction are “missing”, we proposed to use multiple imputation (MI) to fill-in the missing manually-corrected MRI data for measures of normalized whole brain volume (brain parenchymal fraction—BPF) and T2 hyperintense lesion volume (T2LV). Automated and manually corrected MRI measures from 1300 patients enrolled in the Comprehensive Longitudinal Investigation of Multiple Sclerosis at the Brigham and Women's Hospital (CLIMB) were identified. Simulation studies were conducted to assess the performance of MI with missing data both missing completely at random and missing at random. An imputation model including the concurrent automated data as well as clinical and demographic variables explained a high proportion of the variance in the manually corrected BPF (R2 = 0.97) and T2LV (R2 = 0.89), demonstrating the potential to accurately impute the missing data. Further, our results demonstrate that MI allows for the accurate estimation of group differences with little to no bias and with similar precision compared to an analysis with no missing data. We believe that our findings provide important insights for efficient correction of automated MRI measures to obviate the need to perform manual correction on all cases