Prediction of Organ-At-Risk Doses Using An Artificial Intelligence Algorithm: Clinical Validation and Estimated Benefit to Treatment Planning for Lung SBRT
P Brodin1*, L Schulte2, D Pappas2, W Martin3, X Shen3, A Basavatia3, N Ohri1, M Garg1, S Kalnicki3, C Carpenter2, W Tome1, (1) Albert Einstein College of Medicine and Montefiore Medical Center, Bronx, NY, (2) Siris Medical, Newburyport, MA, (3) Montefiore Medical Center, New York, NY
Presentations
SU-E-TRACK 5-7 (Sunday, 7/25/2021) 3:30 PM - 4:30 PM [Eastern Time (GMT-4)]
Purpose: To validate the clinical performance of an artificial intelligence (AI) algorithm to predict pre-treatment organ-at-risk (OAR) doses for lung SBRT using only the planning CT, target and OAR contours. We also assessed the treatment planning benefit of having access to these dose predictions.
Methods: AI models were trained for dose prediction using 213 patients treated with lung SBRT, using a gradient-boosted regression tree algorithm with regularized feature selection. Twenty patients not included in the training set were used to test OAR dose prediction performance, ten with 5x10Gy and ten with 3x18Gy. We also performed blinded re-planning with access to the OAR dose predictions but without access to clinically delivered plans, with target coverage set to equal that of the delivered plans. Differences between predicted and delivered doses were assessed by root-mean square deviation (RMSD) and OAR doses across all three comparisons were evaluated using one-way ANOVA tests.
Results: Prediction performance was good with ANOVA tests showing no significant difference between prediction, delivered and re-planned OAR doses (All p>0.30). The RMSD was 2.9Gy, 5.3Gy, 4.3Gy and 1.7Gy for max dose to spinal cord, esophagus, heart and trachea for 5x10Gy and respectively 1.7Gy, 1.9Gy, 2.9Gy and 2.6Gy for 3x18Gy. Three patients with doses close to the acceptable 52.5Gy max for great vessels had excellent dose prediction with average 3.4Gy difference. In blinded re-planning, lower OAR doses than clinically delivered were achieved for 6/10, 3/10 and 6/10 patients for spinal cord, heart and esophagus at 3x18Gy, and respectively 5/10, 5/10 and 7/10 patients for 5x10Gy.
Conclusion: Pre-treatment prediction of OAR doses showed good agreement with delivered lung SBRT plans and access to these predictions led to lower OAR doses for several patients when re-planned. This prediction will allow upfront discussions of achievable dose fractionation and important OAR and target dose trade-offs.
Funding Support, Disclosures, and Conflict of Interest: Leslie Shulte, Damon Pappas and Colin Carpenter are employed by SIRIS Medical
Handouts
Keywords
Treatment Planning, Lung, Validation
Taxonomy
Contact Email