EcoSta 2017: Start Registration
View Submission - EcoSta2017
A0684
Title: Automated brain hematoma and edema segmentation of CT scans using non-local spatial clustering and the level-set method Authors:  Wei Tu - University of Alberta (Canada) [presenting]
Linglong Kong - University of Alberta (Canada)
Abstract: Hematoma and edema volume are potential predictors of 30-day mortality rate and functional outcome (degree of disability or dependence in the daily activities after a stroke) of patients with intracerebral hemorrhage (ICH). The manual segmentation of hematoma and edema from computed tomography (CT) scans is the common practice, but also a time-consuming and labor-intensive task. The automated segmentation of hematoma and edema is an appealing alternative, but a challenging task due to the poorly defined boundary of edema and the surrounding healthy brain tissue. There is only limited literature on this problem. We propose a novel framework to fill this gap between the theoretical development of segmentation methods and this practical need. Our framework is fully automated and works in an unsupervised fashion. The method uses non-local regularized spatial fuzzy C-means clustering in the initialization stage and level set method in the refinement stage. To evaluate the proposed method, 30 subjects with different size, shape and location of hematoma and edema were used. Compared with manual segmentation results of two independent raters, our method shows an excellent matching in hematoma with an average dice score coefficient 0.92, which is significantly better than the previous methods.