A Hybrid Method for Haemorrhage Segmentation in Trauma Brain CT

Traumatic brain injuries are important causes of disability and death. Physicians use CT or MRI images to observe the trauma and measure its severity for diagnosis and treatment. Due to the overlap of haemorrhage and normal brain tissues, segmentation methods sometimes lead to false results. In this paper, we present a hybrid method to segment the haemorrhage region in trauma brain CT images. Firstly, the images are partitioned to small segments called superpixels and supervoxels in 2D and 3D spaces, respectively. Then the haemorrhage superpixels/supervoxels are grouped using their average intensity as feature. Finally, a distance regularized level-set is used to accurately delineate the exact boundary of the haemorrhage region. Evaluation is performed using the Jaccard overlap measure of our proposed technique against a modified distance regularized level-set and against the manually segmented ground truth. Our results suggest that performing level-set after superpixel/supervoxel segmentation provides better segmentation than superpixel/supervoxel intensity grouping alone and both these schemes perform better than the modified distance regularized level-set evolution method.