Robust and accurate cerebral hemisphere segmentation in non-contrast CT using a 2.5D Dense U-Net

A method is presented for automatically segmenting the cerebral hemi-spheres in non-contrast CT images. The method consists of a convolutional neural network (CNN) that fuses U-Net and DenseNet paradigms into one ar-chitecture. Eight full patient scans were used for training the network. Rotation, shifting and mirroring were used as augmentation steps to enrich the training data set. The method was evaluated on a clinical dataset of 169 patients that received a non-contrast CT. This is a dataset with anatomical and pathologi-cal differences that represents the population of patients that are seen in daily practice. For every patient, three arbitrary orthogonal slices were selected and manually annotated, resulting in 507 reference slices. The method produced few erroneous segmentations and an achieved overall Dice coefficient of 0.944 +- 0.086. The method has proven to be accurate and robust for the segmentation of left and right hemispheres even in the presence of pathology and imaging artifacts.