Semi-automatic liver tumor segmentation with adaptive region growing and graph cuts. (July 2021)

• Record Type: Journal Article
• Authors: Yang, Zhen; Zhao, Yu-qian; Liao, Miao; Di, Shuan-hu; Zeng, Ye-zhan
• Is Part Of: Biomedical signal processing and control. Volume 68(2021)
• Publication Date: 2021-07
• Subjects: Tumor segmentation -- Adaptive region growing -- Graph cuts -- Nonlinear mapping -- CT image; Signal processing -- Periodicals; Biomedical engineering -- Periodicals; Signal Processing, Computer-Assisted -- Periodicals; Image Processing, Computer-Assisted -- Periodicals; Biomedical Engineering -- Periodicals; 610.28
• DOI: 10.1016/j.bspc.2021.102670
• Languages: English
• ISSNs: 1746-8094

Abstract

Highlights: An adaptive region growing method based on Kullback-Leibler divergence is proposed. Graph cuts combined with nonlinear mapping are developed to segment liver tumors. The method can segment the tumors with low contrast and weak boundary precisely. It does not require a heavy training process or a pre-segmentation of liver region. It can also be extended to some other organ or tissue segmentation tasks. Abstract: Segmenting liver tumors from computed tomography (CT) images plays a very important role in computer-aided diagnosis, surgical planning, and treatment monitoring. However, accurate and robust segmentation of the tumors remains a challenging issue, due to low contrast and vague boundaries between the tumors and surrounding tissues as well as the wide variations of the tumors in intensity, shape, and location across patients. In this paper, we developed an effective method for liver tumor segmentation with adaptive region growing and graph cuts. First, initial segmentation results for liver tumors and the regions of interest (ROIs) that contain the tumors are extracted by adaptive region growing with a manual selected seed specified for each tumor region. Then, the ROIs are enhanced by Gaussian fitting based nonlinear mapping according to the intensity distributions of the initially segmented tumor regions. Finally, the enhanced information combined with gradient information is integrated into graph cuts to extract the tumors from the ROIs effectively and accurately. The method is non-sensitive to noise and does not involve a pre-segmentation of liver or a complicated and tedious procedure of training. Results on 3Dircadb dataset demonstrate that the method achieves much better comprehensive performance on liver tumor segmentation compared with many art-of-state methods and has a huge advantage in segmenting the tumors with low contrast, small size, and weak boundaries.