Segmentation of liver anatomy and pathology

Minimally invasive therapy needs the creation of an innovative augmented reality system enabling fast segmentation of targeted organs and the classification of their tissues. In the contest of the ARIS*ER project, this paper presents a new framework to segment liver, vessels and tumors using MRI and CT images. First a coarse to fine approach is used to delineate the surface liver. Then, using this result as a mask on the original data, we implemented a 3D automatic clustering method to classify parenchyma, vessel and tumor voxels. In the same time, we investigated a tube and blob filtering method in order to enhance tumors and vessels in original data. This paper describes these techniques and proposes a study of the tests made on phantom and real patients' data. The results are focused on the quality of the segmentation and the processing time. The segmentation is compared to a hand made segmentation taken as a gold standard for characterizing the quality, while we use algorithms reported in the literature to characterize the time. A Dice similarity coefficient superior at 0.93 shows that our algorithm produces robust and efficient liver segmentations. The processing time of 10s/slice is inferior at other times found on the literature and fits the constraints of pre-planning and quality check. Some improvement can be made for the liver surface extraction but our future effort will be put on the vessels and tumors extraction, the results of which are not enough consistent from one dataset to an other.