Program Information
Hierarchical Atlas-Based Segmentation of the Human Skeleton in CT Images
Y Fu , S Liu , H Li , D Yang*, Washington University School of Medicine, St. Louis, MO
Presentations
WE-G-201-2 (Wednesday, August 2, 2017) 4:30 PM - 6:00 PM Room: 201
Purpose: Proper segment and registration of patient skeleton could facilitate direct comparison and correlation analysis among the pre-, in- and post-treatment images. Image noises, artifacts and partial volume effect lead to unclear bone boundaries and difficulties in separating bones in close proximities, e.g. the vertebrae. Low intensity in spongy and softer bones (e.g. bone marrow, ribs) poses additional difficulties in separating bony structures from soft tissues. To automatically segment and identify 62 different pieces of bones of the human skeleton from CT, an atlas-based hierarchical segmentation framework is proposed.
Methods: An atlas consists of 62 pieces of manually contoured bones was constructed using the female whole body CT dataset from the Visible Human Project. The skeleton of an input patient dataset was roughly segmented as one whole piece using hysteresis thresholding. The skeleton was aligned with the atlas skeleton using Iterative Closest Point (ICP) to initialize the subsequent image registration. The alignment of each individual bone was further optimized locally to improve the alignment accuracy. A kinematic model of human shoulder and pelvic joints was built to support patient body and limb’s postural variability. Each individual bone was segmented by traversing a hierarchical anatomical tree using Morphon deformable image registration.
Results: The proposed method was evaluated on multiple patient datasets, including 14 partial skeleton and 5 whole skeleton datasets. The overall average Dice coefficient is 0.90. The average PSEs are 0.41 mm for mandible, 0.62 mm for cervical vertebrae, 0.92 mm for thoracic vertebrae, and 1.45 mm for pelvis bones, which are better or comparable to other recently published studies.
Conclusion: An image processing framework is proposed to automatically and accurately segment 62 individual bones of the human skeleton in the CT images. The proposed method can be useful in many clinical and research applications in multiple medical disciplines.
Funding Support, Disclosures, and Conflict of Interest: Funding: AHRQ R01-HS022888 No conflict of interest Disclosures: Authors have technology licensing fee from Viewray
Contact Email: