Manual-based and atlas-based segmentation

Brain masks for the pre-operative MMC group were resampled from their corresponding post-operative counterparts after affine alignment. Automatic segmentation of the unmyelinated white matter, ventricles and cerebellum was thereafter performed using template preterm neonatal brain and deep learning-based segmentations which provided good initial segmentation quality [20–22]. Unmyelinated white matter is defined as the brain parenchyma underneath the cortical plate and above the basal ganglia. We use this definition to include any other transitory cerebral structures which may be evident in this region, such as the subplate. Additionally, our automatic segmentations initially included all intracranial cerebro-spinal fluid (CSF). This was thereafter manually corrected to include only the ventricular system which included the lateral, third and fourth ventricle; the cerebral aqueduct; cavum septum pellucidum (CSP) and the cavum vergae (CV). In our MMC cohort, as demonstrated in previous literature, we have observed several cases where the CSP and CV were absent, and therefore, it was simpler at this current stage to include them in our ventricular parcellations [23]. All automatically segmented structures were manually corrected by a trained obstetrician (NM) and corrected as necessary thereafter by a consultant board-certified paediatric neuroradiologist (MA) after which meshes were generated using ITK-Snap™ [24]. This process of manual refinement was necessary due to the complexity of intracranial abnormalities seen in MMC, such as direct apposition of the cortex to the skull in Chiari II malformation, and inability to differentiate between structures especially if compressed in a small posterior fossa.