Can AI Auto-Segmentation Affect Treatment Plan Approval Decisions?
M Jones1*, K Mcghee1, P Mavroidis2, S Stathakis4, (1) Mays Cancer Center - MD Anderson Cancer Center, San Antonio, TX, (2) University of North Carolina, Chapel Hill, NC
Presentations
PO-GePV-T-292 (Sunday, 7/10/2022) [Eastern Time (GMT-4)]
ePoster Forums
Purpose: To assess radiotherapy treatment plans quality using AI segmented structures during evaluation. We are interested to study how dosimetric objectives are affected if expert structures were to be replaced by AI ones.
Methods: For this study, we included 97 previously treated Head and Neck Patients. Plans were optimized using Pinnacle or RaySearch by various dosimetrists and initial contours were segmented by two physicians. The patient CTs were anonymized and exported to a commercially available AI system for organ segmentation. The dose distribution from the approved plan was then evaluated using the AI contours and the approved dosimetric objectives.
Results: No statistical significance was observed between the dosimetric objectives when the same treatment plan was evaluated with the two structure sets. It was noted that in general, the spinal cord segmented by the AI software had a smaller diameter than the physician segmented one. Cochleas, brachial plexuses, and optic nerves had large relative differences due to their small volumes. The oral cavities had the largest difference because the expert-based ones were segmented smaller due to their preference and had the most variability. However, the oral cavity mean dose was lower on average for the physician segmented ones.
Conclusion: Approved treatment plans were evaluated against AI-generated contours to study if the dosimetric objectives are still met. No significant differences have been observed. More results are necessary to validate the initial ones. Furthermore, a planning study comparison utilizing both structure sets might be necessary under control conditions to further investigate the effect of AI-generated contours on the treatment planning optimization.
Keywords
Not Applicable / None Entered.
Taxonomy
Contact Email