Author: Shari Damast, Svetlana Kuznetsova, Christopher J. Tien 👨🔬
Affiliation: Yale University School of Medicine, Department of Therapeutic Radiology, Yale University School of Medicine 🌍
Purpose: Current contour similarity evaluation approaches (Dice Similarity Coefficient, Mean Distance to Agreement) are limited to geometric agreement without assessment of ultimate dosimetric impact. We introduce a novel voxel-based similarity (VBS) methodology which incorporates geometric and dosimetric information. This methodology was applied to evaluate autocontouring software.
Methods: The VBS was utilized to assess similarity for Organs at Risk (OAR) contours against attending physician reference contours for 15 High-Dose-Rate (HDR) single-channel intravaginal brachytherapy patients with applicator in situ on the CT image set. Two commercial systems were used to autocontour bladder, rectum, sigmoid, and bowel structures for the 15 HDR cases. Doses from each voxel of the autocontours were individually scored: if an autocontour voxel was outside of the reference structure then it was assigned a positive dose value (VBS=Di) and any voxels of a reference structure that were not included in a generated contour were assigned a negative corresponding dose value (VBS=-Di). If an autocontour voxel was within the reference structure, the VBS=0. Histograms for each structure provided direct visualization of dosimetric effect resulting from the geometric differences, and helps quantify as to whether each OAR was being “over-contoured” or “under-contoured.” Extracted histogram quantities included 95th percentile, and the fraction of voxels of a generated structure with a 0-scoring. The correlation of these metrics against traditionally used DSC were evaluated along with a dosimetric D2cc metric using Pearson correlation coefficient.
Results: A positive correlation was determined between DSC and 0-scoring fraction for the bladder, rectum, sigmoid, and bowel (0.95, 0.95, 0.97, 0.82). The correlations between the 95th percentile and D2cc were 0.86, 0.86, 0.60, and 0.79 for the same OARs respectively.
Conclusion: The new VBS approach not only incorporates geometric information, but also quantifies dosimetric impact. The VBS showed moderate to high correlation with traditional approaches for autocontouring performance.