Automatic segmentation of solitary pulmonary nodules based on local intensity structure analysis and 3D neighborhood features in 3D chest CT images

Bin Chen; Takayuki Kitasaka; Hirotoshi Honma M.D.; Hirotsugu Takabatake M.D.; Masaki Mori M.D.; Hiroshi Natori M.D.; Kensaku Mori

doi:10.1117/12.911782

Translator Disclaimer

23 February 2012 Automatic segmentation of solitary pulmonary nodules based on local intensity structure analysis and 3D neighborhood features in 3D chest CT images

Bin Chen, Takayuki Kitasaka, Hirotoshi Honma M.D., Hirotsugu Takabatake M.D., Masaki Mori M.D., Hiroshi Natori M.D., Kensaku Mori

Author Affiliations +

Proceedings Volume 8315, Medical Imaging 2012: Computer-Aided Diagnosis; 831538 (2012) https://doi.org/10.1117/12.911782
Event: SPIE Medical Imaging, 2012, San Diego, California, United States

Abstract

This paper presents a solitary pulmonary nodule (SPN) segmentation method based on local intensity structure analysis and neighborhood feature analysis in chest CT images. Automated segmentation of SPNs is desirable for a chest computer-aided detection/diagnosis (CAS) system since a SPN may indicate early stage of lung cancer. Due to the similar intensities of SPNs and other chest structures such as blood vessels, many false positives (FPs) are generated by nodule detection methods. To reduce such FPs, we introduce two features that analyze the relation between each segmented nodule candidate and it neighborhood region. The proposed method utilizes a blob-like structure enhancement (BSE) filter based on Hessian analysis to augment the blob-like structures as initial nodule candidates. Then a fine segmentation is performed to segment much more accurate region of each nodule candidate. FP reduction is mainly addressed by investigating two neighborhood features based on volume ratio and eigenvector of Hessian that are calculates from the neighborhood region of each nodule candidate. We evaluated the proposed method by using 40 chest CT images, include 20 standard-dose CT images that we randomly chosen from a local database and 20 low-dose CT images that were randomly chosen from a public database: LIDC. The experimental results revealed that the average TP rate of proposed method was 93.6% with 12.3 FPs/case.

Citation Download Citation

Bin Chen, Takayuki Kitasaka, Hirotoshi Honma M.D., Hirotsugu Takabatake M.D., Masaki Mori M.D., Hiroshi Natori M.D., and Kensaku Mori "Automatic segmentation of solitary pulmonary nodules based on local intensity structure analysis and 3D neighborhood features in 3D chest CT images", Proc. SPIE 8315, Medical Imaging 2012: Computer-Aided Diagnosis, 831538 (23 February 2012); https://doi.org/10.1117/12.911782

ACCESS THE FULL ARTICLE