Room: Exhibit Hall | Forum 6
Purpose: We evaluated a Fully Convolutional Network (FCN) autosegmentation model on breath-hold CBCT scans and compared its accuracy with human expert inter-observer variability in contouring abdominal organs at risk (OAR).
Methods: Five radiotherapy patients were imaged under breath-hold on a high-speed ring gantry kV CBCT system (HalcyonTM, Varian Medical Systems, Palo Alto, CA). Three dosimetrists, one radiation oncologist and one medical analyst independently delineated seven abdominal OARs in each scan. A sixth set of contours was automatically generated by an independently-trained FCN model. Consensus contours for each OAR were computed via the STAPLE algorithm, from which average symmetric surface distance (ASSD) and Dice similarity coefficient (DSC) were calculated. A Mann-Whitney U test was performed to assess the statistical significance of the difference between model and expert performance.
Results: The mean ASSDs across all five CBCT scans for the experts was compared to that of the model for the liver (0.80 mm vs 1.15 mm), pancreas (2.70 mm vs 2.33 mm), spleen (0.69 mm vs 1.09 mm), stomach-duodenum (1.72 mm vs 2.27 mm), right kidney (0.87 mm vs 1.67 mm) and left kidney (0.87 mm vs 0.96 mm), respectively. The corresponding average DSCs were (0.98 vs 0.97), (0.77 vs 0.84), (0.96 vs 0.94), (0.91 vs 0.89), (0.96 vs 0.91) and (0.96 vs 0.95). The difference in DSC between the human experts and the model was not found to be statistically significant (p < 0.025) for all structures.
Conclusion: This work shows that excellent inter-observer contouring agreement is attainable in breath-hold abdominal CBCT scans acquired on a high-speed ring gantry system. We also demonstrate that a FCN model is a promising approach for automatic delineation of these OARs, which is required for CBCT-guided online adaptive radiotherapy.
Funding Support, Disclosures, and Conflict of Interest: Both authors are employed by Varian Medical Systems, Palo Alto, CA. This work was funded by NIH # U01EB023822.