| dc.description.abstract | Detection of brain development in fetal and neonatal MRI is reliant on the segmentation of the brain in different tissue classes, however since this is cumbersome and time-consuming automatization could simplify the process.
The segmentation of these brains is difficult due to many reasons such as motion artefacts and intensity inhomogenities, as such we aim to test two convolutional neural networks on multiple datasets to determine the most effective one for brain extraction and segmentation.
The two methods used are a UNet and a UNet++ based Convolutional Neural Network to segment the brain into 7 brain classes: intracranial volume, gray matter, white matter, ventricles, cerebellum, deep grey matter, brainstem and spinal cord.
To evaluate the performance of these two methods, the Dice coefficient (DC) and mean surface distance (MSD) per tissue class were computed between expert
manual and automatic annotations. Dice scores where varying being able to at best compete with my predecessor and at worst performing far worse, however mean-surfance distances were able to out-perform khalili et al. | |
