Abstract
Purpose: Effectively quantifying metastatic tumour involvement in the spine requires accurate vertebral segmentation. Automated techniques such as thresholding or region growing have difficulty defining boundaries between tumour tissue and surrounding soft tissue if lytic disease breaches the vertebral cortical shell. It is hypothesized that the application of image registration techniques may afford a potential solution to automating segmentation of metastically-involved vertebrae with cortical shell destruction. The objective of this study is to validate deformable registration as a means to automate the segmentation of tumour-bearing vertebrae through the transformation of atlas segmentations.
Methods: CT scans were collected from 6 patients (T4-L5) with spinal metastases secondary to breast cancer. Healthy levels from the patients were cropped and segmented using a combination of thresholding and manual delineation (Amira 3.1.1, TGS Berlin) to obtain the atlas for each vertebral level. After spatial alignment, metastatically involved vertebral levels were segmented by a registration of the atlas scan by automated affine registration (Amira) and refined by demons deformable registration (ITK, NLM Bethesda). The algorithm was tested through comparison of 10 vertebral bodies (thoracic and lumbar) segmented using the automated approach against a gold standard segmentation produced by semi-manual thresholding. The quality of the automatic segmentation was determined by calculating how many voxels were concurrently within both automatic and manual segmentation of the scan.
Results: Deformable registration successfully segmented metastatically involved vertebrae with and without breach of the cortical shell. Similar performance was evident when using an atlas from an adjacent level as compared to using an atlas of the identical vertebral level. Quality of the automatic segmentation ranged from 87.67%–96.22% concurrency. Comparisons of inter-user semi-manual segmentations yielded a similar maximum of 96% concurrency. Analysis speed was 10 to 15 times faster using the automated technique.
Conclusions: By maintaining the atlas morphology, atlas-based segmentations are able to accurately differentiate between trans-cortical tumours and surrounding soft tissue, overcoming problems inherent to more conventional automated segmentation techniques. Clinical application of this segmentation algorithm centers on tumour quantification and tracking progression of treatment effect and metastatic disease pathology. Funding: Other Education Grant Funding Parties: Canadian Breast Cancer Research Alliance, Sunnybrook & Women`s College Research Institute
Correspondence should be addressed to Cynthia Vezina, Communications Manager, COA, 4150-360 Ste. Catherine St. West, Westmount, QC H3Z 2Y5, Canada
