An automated method for accurate vessel segmentation. (6th April 2017)

• Record Type: Journal Article
• Authors: Yang, Xin; Liu, Chaoyue; Le Minh, Hung; Wang, Zhiwei; Chien, Aichi; Cheng, Kwang-Ting (Tim)
• Is Part Of: Physics in medicine & biology. Volume 62:Number 9(2017:May)
• Page Start: 3757
• Page End: 3778
• Publication Date: 2017-04-06
• Subjects: vessel segmentation -- signal-to-noise ratio -- contrast enhancement -- circle of Wills -- retinal vessels -- aneurysm development analysis; Biophysics -- Periodicals; Medical physics -- Periodicals; 610.153
• DOI: 10.1088/1361-6560/aa6418
• Languages: English
• ISSNs: 0031-9155

Abstract

Abstract: Vessel segmentation is a critical task for various medical applications, such as diagnosis assistance of diabetic retinopathy, quantification of cerebral aneurysm's growth, and guiding surgery in neurosurgical procedures. Despite technology advances in image segmentation, existing methods still suffer from low accuracy for vessel segmentation in the two challenging while common scenarios in clinical usage: (1) regions with a low signal-to-noise-ratio (SNR), and (2) at vessel boundaries disturbed by adjacent non-vessel pixels. In this paper, we present an automated system which can achieve highly accurate vessel segmentation for both 2D and 3D images even under these challenging scenarios. Three key contributions achieved by our system are: (1) a progressive contrast enhancement method to adaptively enhance contrast of challenging pixels that were otherwise indistinguishable, (2) a boundary refinement method to effectively improve segmentation accuracy at vessel borders based on Canny edge detection, and (3) a content-aware region-of-interests (ROI) adjustment method to automatically determine the locations and sizes of ROIs which contain ambiguous pixels and demand further verification. Extensive evaluation of our method is conducted on both 2D and 3D datasets. On a public 2D retinal dataset (named DRIVE (Staal 2004 IEEE Trans. Med. Imaging 23 501 –9 )) and our 2D clinical cerebral dataset, our approach achieves superior performance to the state-of-the-art methods including a vesselness based method (Frangi 1998 Int. Conf. on Medical Image Computing and Computer-Assisted Intervention ) and an optimally oriented flux (OOF) based method (Law and Chung 2008 European Conf. on Computer Vision ). An evaluation on 11 clinical 3D CTA cerebral datasets shows that our method can achieve 94% average accuracy with respect to the manual segmentation reference, which is 23% to 33% better than the five baseline methods (Yushkevich 2006 Neuroimage 31 1116 –28 ; Law and Chung 2008 European Conf. on Computer Vision ; Law and Chung 2009 IEEE Trans. Image Process .18 596 –612 ; Wang 2015 J. Neurosci. Methods 241 30 –6 ) with manually optimized parameters. Our system has also been applied clinically for cerebral aneurysm development analysis. Experimental results on 10 patients' data, with two 3D CT scans per patient, show that our system's automatic diagnosis outcomes are consistent with clinicians' manual measurements.