Automatic contouring of brachial plexus using a multi-atlas approach for lung cancer radiation therapy

Jinzhong Yang; Arya Amini; Ryan Williamson; Lifei Zhang; Yongbin Zhang; Ritsuko Komaki; Zhongxing Liao; James Cox; James Welsh; Laurence Court; Lei Dong

doi:10.1016/j.prro.2013.01.002

Automatic contouring of brachial plexus using a multi-atlas approach for lung cancer radiation therapy

Pract Radiat Oncol. 2013 Oct-Dec;3(4):e139-47. doi: 10.1016/j.prro.2013.01.002. Epub 2013 Feb 9.

Authors

Jinzhong Yang¹, Arya Amini², Ryan Williamson³, Lifei Zhang³, Yongbin Zhang³, Ritsuko Komaki⁴, Zhongxing Liao⁴, James Cox⁴, James Welsh⁴, Laurence Court³, Lei Dong⁵

Affiliations

¹ Department of Radiation Physics, The University of Texas MD Anderson Cancer Center, Houston, Texas. Electronic address: jyang4@mdanderson.org.
² University of California Irvine School of Medicine, Irvine, California; Department of Radiation Oncology, The University of Texas MD Anderson Cancer Center, Houston, Texas.
³ Department of Radiation Physics, The University of Texas MD Anderson Cancer Center, Houston, Texas.
⁴ Department of Radiation Oncology, The University of Texas MD Anderson Cancer Center, Houston, Texas.
⁵ Department of Radiation Physics, The University of Texas MD Anderson Cancer Center, Houston, Texas; Scripps Proton Therapy Center, San Diego, California.

PMID: 24674411
DOI: 10.1016/j.prro.2013.01.002

Abstract

Purpose: To demonstrate a multi-atlas segmentation approach to facilitating accurate and consistent delineation of low-contrast brachial plexuses on computed tomographic images for lung cancer radiation therapy.

Methods and materials: We retrospectively identified 90 lung cancer patients with treatment volumes near the brachial plexus. Ten representative patients were selected to form an atlas group, and their brachial plexuses were delineated manually. We used deformable image registration to map each atlas brachial plexus to the remaining 80 patients. In each patient, a composite contour was created from 10 individual segmentations using the simultaneous truth and performance level estimation algorithm. This auto-delineated contour was reviewed and modified appropriately for each patient. We also performed 10 leave-one-out tests using the 10 atlases to validate the segmentation accuracy and demonstrate the contouring consistency using multi-atlas segmentation.

Results: The multi-atlas segmentation took less than 2 minutes to complete. Contour modification took 5 minutes compared with 20 minutes for manual contouring from scratch. The multi-atlas segmentation from the 10 leave-one-out tests had a mean 3-dimensional (3D) volume overlap of 59.2% ± 8.2% and a mean 3D surface distance of 2.4 mm ± 0.5 mm. The distances between the individual and average contours in the 10 leave-one-out tests demonstrated much better contouring consistency for modified contours than for manual contours. The auto-segmented contours did not require substantial modification, demonstrated by the good agreement between the modified and auto-segmented contours in the 80 patients. Dose volume histograms of auto-segmented and modified contours were also in good agreement, showing that editing auto-segmented contours is clinically acceptable in view of the dosimetric impact.

Conclusions: Multi-atlas segmentation greatly reduced contouring time and improved contouring consistency. Editing auto-segmented contours to delineate the brachial plexus proved to be a better clinical practice than manually contouring from scratch.