Room: ePoster Forums
Purpose: Patients often present to radiation oncology with recent diagnostic computed tomography (CT) scans. Treatment planning scans that are acquired for dose calculation and positioning provide duplicate information (to the diagnostic CT). If a segmentation algorithm could consistently identify tissue types across different scanners, this tool could be used to generate synthetic CTs based on the most likely values of tissue types determined from a well characterized known scanner. This may allow for accurate dose calculation on diagnostic CTs, offering cost-effective image reduction.
Methods: Twelve chest CT scans acquired on a single scanner were segmented with Fuzzy c-means clustering using a custom Matlab script with variable numbers of clusters and optimization parameters. Synthetic CTs were created by mapping each voxel to the mean centroid value of its cluster averaged over all patients. Segmentation accuracy was evaluated visually and by comparing cluster centroids between scans to determine if the clusters were centering on the same tissue type in each scan.
Results: Twelve clusters provided the most accurate segmentation (6-16 clusters were evaluated). Visual examination showed that the lowest (lung) and highest (bone) CT number clusters were correctly identified on each patient. There was some variance in tissue classification in the fat range. The centroid mean values over all patients ranged from -943 â€“ 770 HU. The standard deviations of the centroid means over all patients ranged from 17 â€“ 171 HU. 72% of the voxels were in clusters with mean centroids -142 â€“ 105 HU, which had standard deviations from 17 â€“ 69 HU (no other cluster contained > 5% of the voxels). This level of variance is expected to yield << 5% dose inaccuracy.
Conclusion: It is feasible to identify tissue type by clustering with sufficient accuracy to calculate dose. Next steps are to verify these results across multiple scanners.