Towards Contour Quality Assurance of Cardiac Structures with Automatic Segmentation
C Uche1, H Geng2, J Yu3, E Gore4, Y Xiao5, (1) University of Pennsylvania, Philadelphia, PA, (2) University Of Pennsylvania, Philadelphia, PA,(3) NRG Oncology Statistics and Data Management Center, Philadelphia, PA,(4) Medical College of Wisconsin, Milwaukee, Wisconsin,(5) University of Pennsylvania, Philadelphia, PA
Presentations
WE-D-TRACK 1-1 (Wednesday, 7/28/2021) 2:00 PM - 3:00 PM [Eastern Time (GMT-4)]
Purpose: This study aims to determine the clinical readiness of two auto-segmentation approaches for cardiac structures contouring quality assurance.
Methods: Five experienced radiation oncologists delineated the pericardium, atria, and ventricles of 469 RTOG0617 CT scans using a consistent guideline. We developed atlas-based automatic contouring (ABAC) library using 100 randomly sampled CT scans. The CT scans of 200 patients (including the 100 scans for ABAC) were trained on an in-housed developed convolutional neural network (CNN) model. The remaining patients were used to assess the performance of the two models, based on the Dice similarity coefficient (DSC), Hausdorff distances (HD), and mean distance-to-agreement (MDA). For dosimetric comparison, the mean dose (Dmean), and the maximum dose (Dmax), were computed for all the cardiac structures. Relative dose-volume factors, including V5 - V60, were analyzed to study the effect of dose heterogeneity. Spearman correlation enabled us to evaluate the relationships between geometric and dosimetric metrics.
Results: Compared with ABAC, the geometric metrics (DSC, HD, and MDA) of CNN for the cardiac structures depicted a better alignment with the physician-drawn contours. Auto-segmentation with CNN also attained relatively lower Dmean, Dmax, and V5 to V60 to the cardiac structures. However, significant mean and maximum dose differences between the two auto-segmentation methods in the atria and ventricles highlighted a non-negligible disparity between the two auto-segmentation approaches. Analysis of the geometric and dosimetric metrics showed no clear relationship with the automatic contours due to dose heterogeneity.
Conclusion: Although auto-segmentation shows promise to reduce cardiac structures’ contouring inconsistencies, the heterogeneous distribution of radiation doses within the heart segments negates the use of geometric metrics as suitable surrogates for dosimetric parameters. Thus, auto-segmentation would continue to require a thorough validation of a qualified expert to ensure its safe clinical use.
Handouts
Keywords
Contour Extraction, Quality Assurance, Segmentation
Taxonomy
IM/TH- image Segmentation: General (Most aspects)
Contact Email