Automating Quantitative Measures of an Established Conventional MRI Scoring System for Preterm-Born Infants Scanned between 29 and 47 Weeks' Postmenstrual Age.

Background And Purpose: Conventional MR imaging scoring is a valuable tool for risk stratification and prognostication of outcomes, but manual scoring is time-consuming, operator-dependent, and requires high-level expertise. This study aimed to automate the regional measurements of an established brain MR imaging scoring system for preterm neonates scanned between 29 and 47 weeks' postmenstrual age.

Materials And Methods: This study used T2WI from the longitudinal Prediction of PREterm Motor Outcomes cohort study and the developing Human Connectome Project. Measures of biparietal width, interhemispheric distance, callosal thickness, transcerebellar diameter, lateral ventricular diameter, and deep gray matter area were extracted manually (Prediction of PREterm Motor Outcomes study only) and automatically. Scans with poor quality, failure of automated analysis, or severe pathology were excluded. Agreement, reliability, and associations between manual and automated measures were assessed and compared against statistics for manual measures. Associations between measures with postmenstrual age, gestational age at birth, and birth weight were examined (Pearson correlation) in both cohorts.

Results: A total of 652 MRIs (86%) were suitable for analysis. Automated measures showed good-to-excellent agreement and good reliability with manual measures, except for interhemispheric distance at early MR imaging (scanned between 29 and 35 weeks, postmenstrual age; in line with poor manual reliability) and callosal thickness measures. All measures were positively associated with postmenstrual age ( = 0.11-0.94; = 0.01-0.89). Negative and positive associations were found with gestational age at birth ( = -0.26-0.71; = 0.05-0.52) and birth weight ( = -0.25-0.75; = 0.06-0.56). Automated measures were successfully extracted for 80%-99% of suitable scans.

Conclusions: Measures of brain injury and impaired brain growth can be automatically extracted from neonatal MR imaging, which could assist with clinical reporting.