Abstract
DIR algorithms differ by physical principles employed in their design which in turn determines complexity of the transformation that the algorithm allows. Indiscriminate application of different algorithms without regard for their limitations may lead to significant discrepancies in research results and in clinical procedures. We compare two different algorithms in Head and Neck (H&N) patients to assess what aspects of DIR use are most sensitive to differences between algorithms.
H&N patients are enrolled in a protocol which acquires weekly FBCT and daily double CBCT. Each FBCT study is contoured by the same physician. Two DIR algorithms are compared: Small Deformation Inverse-Consistent Linear Elastic (SICLE), and the ITK Diffeomorphic Demons (ITKDD) as implemented in the ITK package. Both algorithms employ significantly different physical principles in their design and consequently impose different restrictions on the complexity of transformation they allow. We compare Jacobian Volume Histograms (JVH), Spatial Discrepancy Volume Histograms (SDVH), BED and physical dose accumulation results and resulting plan evaluation indices.
Analysis of mean Jacobian shows that both algorithms are able to detect changes in structure volumes, though they differ quantitatively from one another and from the ground truth as established by the analysis of changes in contours. Width of Jacobian distributions is very different indicating that Jacobian should not be used as a measure of volume change at a voxel level without independent validation. Analysis of SDVHs shows that dose lookup points implied by both algorithms are separated by 5mm - 10mm over approximately 30% of most volumes. These differences translate into clearly visible though not very significant differences in BED and dose accumulation. We further observe that physical dose accumulation in external beam H&N patients is a good proxy for direct BED accumulation.
Different DIR algorithms may have to be applied selectively in different areas of treatment planning. Acknowledgments: Supported by NIH Grant P01 CA11602.
