Modelling brain development to detect w f to matter injury in term and preterm born neonates

Premature birth occurs during a period of rapid brain growth. In this context, interpreting clinical neuroimaging can be complicatedby the typical changes in brain contrast, size and gyrification occurring in the background to any pathology. To model anddescribe this evolving background in brain shape and contrast, we used a Bayesian regression technique, Gaussian process regression,adapted to multiple correlated outputs. Using MRI, we simultaneously estimated brain tissue intensity on T-1- and T(2)weightedscans as well as local tissue shape in a large cohort of 408 neonates scanned cross-sectionally across the perinatalperiod. The resulting model provided a continuous estimate of brain shape and intensity, appropriate to age at scan, degree ofprematurity and sex. Next, we investigated the clinical utility of this model to detect focal white matter injury. In individualneonates, we calculated deviations of a neonate's observed MRI from that predicted by the model to detect punctate white matterlesions with very good accuracy (area under the curve 4 0.95). To investigate longitudinal consistency of the model, we calculated model deviations in 46 neonates who were scanned on a second occasion. These infants' voxelwise deviations from the modelcould be used to identify them from the other 408 images in 83% (T-2-weighted) and 76% (T-1-weighted) of cases, indicating ananatomical fingerprint. Our approach provides accurate estimates of non-linear changes in brain tissue intensity and shape withclear potential for radiological use.