Hydrocephalus is a disease where an excess of cerebrospinal fluid (CSF) is built up in the brain. It affects approximately 180.000 infants per year in sub-Saharan Africa. Magnetic resonance imaging (MRI) is an advantageous imaging method to diagnose hydrocephalus and examine the amount of fluid in the brain for treatment. Unfortunately, in sub-Saharan Africa there is limited access to MRI scanners. That is why an inexpensive, portable, low-field MRI scanner is being built for the treatment of hydrocephalus in Uganda. One main restriction of this scanner is simplicity in use. Therefore, the goal is to make software for the MRI scanner that automatically makes and processes the scans. Part of the processing is the automatic segmentation of the scan into CSF and brain tissue regions. Automatic segmentation is complex due to noise and artifacts present in low-field scans. Also, automation of segmentation processes is complicated. Therefore, in this thesis project we aimed to realize the foundations for a fast, practical and automatic 3D segmentation method for brain scans obtained by the low-field MRI scanner. First, analytic segmentation methods were investigated. Multiple segmentation methods were applied to different high-field and low-field scans. Data analysis showed that Li’s method, where the intensity non-homogeneity artifact was corrected for during segmentation, improved the segmentation results evidently when the scan was affected by a distinct bias field. The major disadvantage of Li’s method was the number of parameters and initialization values that had to be chosen. Therefore, it would be complex to satisfy the automation requirement by using an analytic segmentation method. However, this led to the idea of integrating Li’s method with a neural network. A neural network would solve the problem of automation, while the incorporation of Li’s method, would lead to the fitting energy of the segmentation being minimized, which could improve the segmentation results. A neural network for the segmentation of CSF, white matter and gray matter and the prediction of the bias field was built. Unfortunately, a low-field dataset was not available to train on. Therefore, for the training of the network, artificial high-field data was used, together with its ground truth segmentations. Then, the network was trained by not only comparing the predictions with the ground truth segmentations and bias field, but also by minimizing the fitting energy of the predicted segmentation. To evaluate the segmentation results Dice scores were computed between the ground truth and the predicted segmentations. The Dice scores of the train and test set showed that the segmentation results of the neural network improved when the analytic segmentation loss was added to the network. To further investigate the promising effect of the analytic segmentation loss, the trained network was transferred to a new dataset containing infant brain scans, for which the ground truth segmentations were ignored. The network was therefore trained on this new dataset by only using the analytic segmentation loss function. Unfortunately, the results of segmentation became worse. This occurred due to non-brain tissues wrongly being segmented in the brain tissue clusters, since the signal intensities were close to each other. After the non-brain tissues were removed by brain masking, the results of segmentation for all tissues had improved. When more low-field brain scans are available the neural network should be transferred to these scans, since automatic segmentation of low-field scans is the final objective. It is recommended to first implement an automatic brain masking technique on the scans for optimal results. The promising results of Li’s method applied to the low-field scans and of transferring the neural network to the infant dataset show excellent future perspective for the fast, practical and automatic 3D segmentation of low-field scans.