PRF-RW: a progressive random forest-based random walk approach for interactive semi-automated pulmonary lobes segmentation

• Published in: International journal of machine learning and cybernetics Vol. 11; no. 10; pp. 2221 - 2235
• Main Authors: Li, Qiang, Chen, Lei, Li, Xiangju, Lv, Xiaofeng, Xia, Shuyue, Kang, Yan
• Format: Journal Article
• Language: English
• Published: Berlin/Heidelberg Springer Berlin Heidelberg 01.10.2020; Springer Nature B.V
• Subjects: Accuracy; Algorithms; Artificial Intelligence; Asthma; Automation; Biopsy; Chronic obstructive pulmonary disease; Complex Systems; Computational Intelligence; Computed tomography; Control; Datasets; Decision trees; Deep learning; Emphysema; Engineering; Image segmentation; Lobes; Lungs; Machine learning; Mechatronics; Medical imaging; Medical screening; Methods; Original Article; Pattern Recognition; Random walk; Robotics; Systems Biology; Watersheds; Random forest; Random walk; Lobes segmentation; Machine learning; Semi-automated segmentation
• ISSN: 1868-8071; 1868-808X
• DOI: 10.1007/s13042-020-01111-9

The computational detection of lung lobes from computed tomography images is a challenging segmentation problem with important respiratory healthcare applications, including emphysema, chronic bronchitis, and asthma. This paper proposes a progressive random forest-based random walk approach for interactive semi-automated pulmonary lobes segmentation. First, our model performs automated segmentation of the lung lobes in a progressive random forest network, eliminating the need for prior segmentation of lungs, vessels, or airways. Then, an interactive lobes segmentation approach based on random walk mechanism is designed for improving auto-segmentation accuracy. Furthermore, we annotate a new dataset which contains 93 scans (57 men, 36 women; age range: 40–90 years) from the Central Hospital Affiliated with Shenyang Medical College (CHASMC). We evaluate the model on our annotated dataset, LIDC ( https://wiki.cancerimagingarchive.net ) and LOLA11 ( http://lolall.com/ ) datasets. The proposed model achieved a Dice score of 0.906 ± 0.106 for LIDC, 0.898 ± 0.113 for LOLA11, and 0.921 ± 0.101 for our dataset. Experimental results show the accuracy of the proposed approach, which consistently improves performance across different datasets by a maximum of 8.2% as compared to baselines model.