Adaptive Predictions of Confidence Regions During Indirect Lung Tumor Tracking
C Remy*1,2, H Bouchard1,2,3, (1) Université de Montréal, Département de Physique, Montréal, QC, CA, (2) Centre de recherche du Centre hospitalier de l'Université de Montréal, Montréal, QC, CA, (3) Département de radio-oncologie, Centre hospitalier de l'Université de Montréal, Montréal, QC, CA.
Presentations
WE-C930-IePD-F2-4 (Wednesday, 7/13/2022) 9:30 AM - 10:00 AM [Eastern Time (GMT-4)]
Exhibit Hall | Forum 2
Purpose: As part of indirect tracking for radiotherapy, the present work addresses the problem of predicting real-time confidence regions along with the future target position while relying exclusively on the surrogate motion. This study examines the ability of the confidence estimation to accurately reflect prediction reliability and to prospectively detect large prediction errors on target positions.
Methods: This work builds on a Bayesian framework for indirect tracking. New adaptive surrogate-based covariance estimates are derived to reflect increasing uncertainty when the breathing conditions differ from those observed during the initial training step. The accuracy of the resulting 95% confidence regions (CR) is assessed on 9 continuous ground truth breathing sequences (simultaneous lung target and two different surrogates) acquired from dynamic MRI on healthy subjects. Evaluation is also conducted on 11 Cyberknife Synchrony treatment fractions of lung tumor motion data. To evaluate the predictive uncertainty's ability to detect large prediction errors, receiver operating characteristic analysis (ROC) is performed.
Results: Using adaptive over constant estimates significantly improves the accuracy of the CR: on average, the proportion of actual target positions lying within the 95% CR is increased by 40 and 35 p.p. with the internal and external surrogates. The adaptive CR expand when the target position cannot be reliably predicted, which corresponds to potentially high prediction errors. The ROC analyses indicate that the proposed uncertainty estimation can detect large prediction errors with high sensitivity (90%) and modest specificity (54% and 47% on average based on internal and external surrogates for the volunteers’ cohort). Preliminary results obtained from Synchrony data suggest matching performance (46% specificity at 90% sensitivity).
Conclusion: Based solely on surrogate monitoring, the proposed framework provides adaptive CR that can be used in detecting large prediction errors. This study opens the way for real-time uncertainty prediction to further assist motion management decisions.
Funding Support, Disclosures, and Conflict of Interest: This research was supported by in part by Elekta, Medteq and NSERC (Collaborative Grant no. CRDPJ/502332-2016) and in part by the Fonds québécois de la recherche sur la nature et les technologies (FRQNT scholarship no. 271036).
Keywords
Bayesian Statistics, Target Localization, Radiation Therapy
Taxonomy
TH- External Beam- Photons: Motion management - intrafraction
Contact Email