Click here to


Are you sure ?

Yes, do it No, cancel

Bayesian Stochastic Frontier Analysis with Missing Data Management as Knowledge-Based Planning for Lung SBRT

A Kroshko1,2*, O Morin3, L Archambault1,2, (1) Universite Laval, Quebec, QC, CA (2) CHU de Quebec - Hotel-Dieu de Quebec, Quebec, QC, CA (3) University of California San Francisco, San Francisco, CA, USA


(Sunday, 7/12/2020)   [Eastern Time (GMT-4)]

Room: AAPM ePoster Library

Purpose: To develop a knowledge-based planning method to predict optimal dosimetric indices for lung SBRT. Most KBP strategies rely on large volume of contoured data, but in clinical practice, only nearby OARs are delineated. To keep KBP clinically relevant, lengthy contouring sessions must be avoided. Thus, we establish a novel method based on multiple imputation to incorporate the data for patients with missing structures in the predictive model.

Methods: A retrospective study of 219 patients, plus 30 for validation, treated for lung SBRT with VMAT technique was made. Prescribed dose to the PTV were 48 or 52 Gy in 4 fractions or 50 Gy in 5 fractions. Bayesian Stochastic Frontier Analysis (BSFA) is used to predict dosimetric indices based on anatomical features with a Markov Chain Monte Carlo algorithm. Geometric parameters were extracted between the PTV and 12 OARs such as, spinal cord, great vessels, esophagus, heart, main bronchus. Minimum distances between the PTV and the OARs for missing structures are estimated as following a truncated normal distribution with an established censored value for its mean specific to each OARs.

Results: For the validation set, mean difference between observed and predicted values are -3.2 ± 2.3 % for V100% for the PTV, 2.4 ± 1.7 Gy for the D0.35cc and 2.1 ± 1.7 Gy for the Dmax of the spinal cord. Furthermore, higher order effects of the minimum distance between PTV and spinal cord were significant parameters to predict the dose for this OAR. No significative changes in the parameter estimates between modification in the censored values and the standard deviation of the distribution demonstrate that the implemented treatment of missing values is robust.

Conclusion: BSFA combined with our implemented missing data imputation proves to be a promising method to predict dosimetric parameters for PTV and OARs for lung SBRT.


Not Applicable / None Entered.


TH- External Beam- Photons: Treatment planning using machine learning/Knowledge Based Planning/automation

Contact Email