Skip to main content

Automatic detection of sclerotic bone metastases in the spine using computed tomography images

Wednesday, October 26, 2011 — Poster Session III

10:00 a.m. – Noon

Natcher Conference Center




  • J Yao
  • J Burns
  • T Wiese
  • R Summers


We introduce a CAD system which detects sclerotic osseous metastases in the spine on computed tomography images. After spine segmentation, a watershed algorithm detects lesion candidates, and 25 quantitative features for each detection are used to train a committee of seven support vector machines (SVMs). A novel feature of our system is the incorporation of a post-watershed graph-cuts merger to address oversegmentation due to noise and extraneous structures, improving program versatility and including more global features than simple thresholding methods permit. The classifier was trained on 12 clinical cases. A performance test was conducted on 10 clinical cases independent of the training set. Ground truth lesions were manually segmented by an expert. The segmentation algorithm detects 72 out of 83 manually segmented lesions with volume greater than 300 mm3. A ten-fold cross-validation results in 71.2% sensitivity ([95% confidence interval (0.631, 0.773)], at an average of 8.76 false positives per case.

back to top