Affiliations: Key Lab of Contemporary Design and Integrated Manufacturing Technology, Ministry of Education, School of Mechanical Engineering, Northwestern Polytechnical University, Xi’an Shaanxi, China
Correspondence:
[*]
Corresponding author: Jia Zheng, Key Lab of Contemporary Design and Integrated Manufacturing Technology, Ministry of Education, School of Mechanical Engineering, Northwestern Polytechnical University, 127 West Youyi Road, Xi’an Shaanxi, 710072, China. E-mail: jiazheng@mail.nwpu.edu.cn.
Abstract: BACKGROUND:Images of industrial cone-beam computed tomography (CBCT) contain noise and beam hardening artifacts, which induce difficulty and low precision in segmenting regions of interest. OBJECTIVE:The primary objective of this study is to improve the segmentation precision of CBCT series slice images. METHODS:This paper presents a method based on the Phansalkar to segment CBCT series slice images precisely. First, the basics of the proposed method and the necessity of changing the local window size are analysed. The adaptive accumulated Phansalkar, which collects each pixel’s classification results in different local windows, is proposed. Second, the bimodal distribution of the histogram is used to calculate the appropriate local window size for each pixel adaptively. Third, the characteristics of the accumulated probability (the accumulated classification results divided by the accumulated times) are analysed, from which an adaptive method is applied to segment the accumulated probability. Last, experiments are conducted on CBCT series slice images of three workpieces and one computer-aided design (CAD) model with internal defects. RESULTS:The proposed new method can segment CBCT images with noise and beam-hardening well. Moreover, for the segmentation of all four CBCT series slice images, the new method acquired the highest BF and AOM scores (1 and 0.9981) with the smallest standard deviation (0.0013) as compared with other existing methods including CMF (continuous max-flow/min cut), MS (mean-shift), DRLSE (distance regularized level set evolution), and ARKFCM (adaptively regularized kernel-based fuzzy c-means clustering). CONCLUSIONS:The experimental results support that our new method can more precisely segment CBCT series slice images with noise and artifacts than many existing methods. Thus, the new method has prospective application value and can provide valuable technical support for the industrial CBCT image post-processing system.
