Automatic MRI-based quantication of brain characteristics in preterm newborns

Abstract

Even though survival of preterm infants has improved in recent years, preterm birth is still associated with developmental impairments, such as cognitive and behavioural problems. Brain development is particularly vulnerable in this population because an important part of development takes place after birth, which could lead to primary brain injury and secondary developmental consequences. Especially the cerebral cortex rapidly develops from a smooth surface to a complexly folded structure in the third trimester of pregnancy and is therefore vulnerable in extra-uterine conditions. Cortical development has also been described to be disturbed by white matter injury. Magnetic resonance imaging (MRI) provides an important non-invasive tool to assess brain development in preterm infants. This thesis describes an automated system for quantification of brain characteristics in preterm newborns based on MR brain images. To allow automatic quantification of brain characteristics based on these images, Chapters 2–5 describe automatic segmentation (i.e. labelling of regions of interest in the image) approaches for MR brain images. Chapter 2 describes an automatic segmentation method for cortical grey matter, unmyelinated white matter and extracerebral cerebrospinal fluid. The method is based on sequential supervised voxel classification and evaluated on coronal MR images of preterm newborns at 30 and 40 weeks postmenstrual age. Chapter 3 evaluates the approach described in Chapter 2 in the segmentation in MR images of ageing adults at an average age of 70 years. Chapter 4 describes an automatic segmentation method that uses a multiscale convolutional neural network to segment neonatal and adult MR brain images into a number of tissue classes. Chapter 5 describes a general medical image segmentation method using a convolutional neural network that is applied to the segmentation of seven tissue classes in MR brain images, the pectoral muscle in MR breast images and the coronary arteries in cardiac CT angiography. Based on automatic segmentations or MR brain images, Chapter 6 describes a system to compute characteristics of the cerebral cortex. The system is evaluated in a cohort of 85 preterm newborns imaged at 30 and 40 weeks postmenstrual age. The descriptors show longitudinal and regional differences, and show a relation with a conventional visual brain abnormality scoring system. Chapter 7 describes cortical morphology in newborns with severe congenital heart disease compared with healthy controls, using the system from Chapter 6. Chapter 8 describes prediction of cognitive and motor performance at later age based on descriptors automatically computed from MR images of preterm newborns, using the segmentation approach from Chapter 4 and the feature computation approach from Chapter 6. The work in this thesis showed that it is possible to automatically compute quantitative descriptors from an MR brain image that are valuable in the assessment of neurodevelopment of preterm infants: from image segmentation, to feature computation, to outcome prediction. Such a system could in the future be implemented in clinical practice to, directly after the image acquisition, compute quantitative measurements and possibly even identify preterm infants at risk of neurodevelopmental impairments.