Fully Automatic 3D Segmentation Measurements of Human Liver Vessels from Contrast-Enhanced CT F. Conversano, E. Casciaro, R. Franchini, and S. Casciaro Institute of Clinical Physiology National Council of Research Lecce, Italy sergio.casciaro@cnr.it A. Lay-Ekuakille Department of Innovation Engineering University of Salento Lecce, Italy Abstract—Aim of the present work was to evaluate the performance of a novel fully automatic algorithm for 3D segmentation and volumetric reconstruction of liver vessel network from contrast-enhanced computed tomography (CECT) datasets acquired during routine clinical activity. Three anonymized CECT datasets were randomly collected and were automatically analyzed by the new vessel segmentation algorithm, whose parameter configuration had been previously optimized on a phantom model. The same datasets were also manually segmented by an experienced operator that was blind with respect to algorithm outcome. Automatic segmentation accuracy was quantitatively assessed for both single 2D slices and 3D reconstruction of the vessel network, accounting manual segmentation results as the reference “ground truth”. Adopted evaluation framework included the following two groups of calculations: 1) for 3D vessel network, sensitivity in vessel detection was quantified as a function of both vessel diameter and vessel order; 2) for vessel images on 2D slices, dice similarity coefficient (DSC), false positive ratio (FPR), false negative ratio (FNR), Bland-Altman plots and Pearson correlation coefficients were used to judge the correctness of single pixel classifications. Automatic segmentation resulted in a 3D vessel detection sensitivity of 100% for vessels larger than 1 mm in diameter, 64.6% for vessels in the range 0.5-1.0 mm and 27.8% for smaller vessels. An average area overlap of 99.1% was obtained between automatically and manually segmented vessel sections, with an average difference of 0.53 mm 2 . The corresponding average values of FPR and FNR were 1.8% and 1.6%, respectively. Therefore, the tested method showed significant robustness and accuracy in automatic extraction of the liver vessel tree from CECT datasets. Although further verification studies on larger patient populations are required, the described algorithm has an exciting potential for supporting liver surgery planning and intraoperative resection guidance. Keywords—biomedical image processing; medical diagnostic imaging; computed tomography imaging; automatic segmentation; liver surgery. I. INTRODUCTION Planning systems for liver surgery are typically based on dedicated algorithms for accurate segmentation of relevant anatomic structures within high resolution images (e.g., computed tomography, CT) [1,2]. Complexity of internal liver anatomy, coupled with the significant variability between different patients and the close proximity of other abdominal organs, makes the automatic liver segmentation a particularly challenging task [3-7]. For this reason, recent literature reported numerous examples of implementation and experimental tests of possible new approaches to liver segmentation, each one characterized by its specific “cost- effectiveness compromise” [8-13]. In this context, our research group has also introduced a fully automatic algorithm for 3D segmentation of liver vessel network from contrast-enhanced CT (CECT) images [1]. The accurate segmentation of liver vessels, in fact, represents an essential pre-requisite for several liver therapies, like surgical removal of cancers, cryoablation, radiofrequency ablation [15- 17]. First experimental validation of our vessel segmentation algorithm was conducted on a phantom that precisely reproduced the vessel network of a human liver [1]. In the referred phantom study, we first performed an optimization of the algorithm parameter values and then we quantitatively assessed its accuracy in fully automatic vessel detection, obtaining the unprecedented sensitivity of 100% for phantom vessels larger than 1 mm in diameter. In the present study we performed the first experimental validation of the above mentioned algorithm on real CECT datasets acquired during routine clinical activity. II. MATERIALS AND METHODS A. Patient Data Acquisition Three anonymized CECT datasets were randomly selected among those collected for a previous study designed for a different purpose [12]. Patients enrolled for the referred study were not recruited on the basis of specific inclusion or exclusion criteria and we kept this approach valid also for the present study, in order to test the effectiveness of our algorithm on random routine cases. The aforementioned study [12] fully respected national privacy laws and had been also approved by the ethics committee. Adopted CT protocols followed the routine 978-1-4799-2921-4/14/$31.00 ©2014 IEEE