Atlas-Free Method of Periventricular Hemorrhage Detection from Preterm Infants’ T1 MR Images

Abstract

Detection of hemorrhages in the periventricular white matter region of infant brains is crucial since if left untreated it causes neuro-developmental deficits in later life. However, noise and motion artefacts are introduced while scanning infant brains due to small brain size and movement during scanning. Furthermore, a vast majority of traditional brain lesion detection algorithms which require accurate segmentation of the white matter region often rely on brain atlases to guide the segmentation. However, reliable brain atlases are hard to obtain for preterm infant brains which undergo rapid structural changes. To address this gap in published literature, we propose a novel method for hemorrhage detection which does not require a brain atlas. Instead of attempting accurate segmentation, the proposed method detects the ventricles and then samples a region of white matter around the ventricles. Based on the normal distribution of intensities in this tissue sample, the outliers are designated as hemorrhages. Heuristics based on size and location of the detected outliers are used to eliminate false positives. Results on an expert-annotated dataset demonstrate the effectiveness of the proposed method.

Supported by CIHR, NeuroDevNet, Alberta Innovates (iCORE) Research Chair program, and NSERC. DICOM slices with marked ground truth for preterm neonates’ periventricular hemorrhage detection provided by Dr. Steven Miller and his team at SickKids Hospital, Toronto, Canada.