2373. Automated Measurement of Vertebral Deformities and Detection of Lumbar Scoliosis in Spine DXA Scans
Authors * Denotes Presenting Author
  1. Abhinav Suri *; University of Pennsylvania
  2. Sisi Tang; University of Pennsylvania
  3. Samantha Turner; University of Pennsylvania
  4. Ashley Terry; University of Pennsylvania
  5. Eusha Hasan; University of Pennsylvania
  6. Andy Chen; University of Pennsylvania
  7. Chamith Rajapakse; University of Pennsylvania
Vertebral fractures/deformities comprise nearly 50% of all fractures in patients diagnosed with osteoporosis. Specifically, lumbar spinal deformities are highly prevalent among older patients with osteoporosis and other degenerative bone diseases. Furthermore, these deformities cause a curvature of the spine (scoliosis), leading to a higher risk of fracture and pain in patients. Currently, DXA scans report vertebrae specific BMD scores; however, they do not report vertebral deformities or the presence of lumbar curvature. Reporting these deformities requires the manual measurement of heights for each vertebral body which is time-consuming. Computer vision algorithms can shorten the time to analyze imaging studies by automatically extracting specific features. The purpose of this study was to determine if such algorithms could output spine metrics (vertebral body heights) that are useful for evaluating vertebral deformities and curvatures in Spine DXA scans. We report the development of an algorithm that can locate vertebral body landmarks for height calculations and find vertebral body centroids for scoliosis evaluation in less than 0.5 seconds for Spine DXA scans.

Materials and Methods:
Spine DXA reports (1282 subjects, age 64.5±10.7yrs, female) were processed to isolate the relevant DXA image. From the DXA scan, the image was thresholded to find an outline of the spine itself, and a Hough Line Transformation was applied to determine planes that separated lumbar bodies L1-L5. Intersection points between the planes and the outline of the spine were calculated for each vertebral body (n=6410) producing four corner points that were used for the measurement of vertebral body lateral heights (right lateral vs left lateral height) and Cobb angles (calculated from the centroid of four corners) in the coronal plane. The algorithm was evaluated using a key-point error distance metric that measures the distance of predicted corner points to manually marked corner points of vertebral bodies.

The algorithm was able to detect relevant corner points rapidly and with high accuracy. On a set of 1000 scans, each scan was processed in an average of 0.035 ± 0.012 seconds. The average key-point error distance was 0.13mm when evaluated on a testing set of 100 manually annotated scans (age 64.3±5.3yrs, female).

The computer vision algorithm was able to determine morphometric measurements for detecting vertebral body corner points with high accuracy on DXA images. Additionally, it was able to do so quickly. These corner points can be used to calculate vertebral deformities and Cobb angles for scoliosis grading. Thus, we were able to show this algorithm is able to automatically extract measurements for rapid quantification of vertebral bone health and potential risk of fracture. This automated approach could simplify the screening, detection of changes, further cross-sectional imaging workup, and surgical planning in patients with vertebral deformities and fractures by reducing the burden on radiologists who have to do these measurements manually.