Room: Karl Dean Ballroom C
Purpose: To develop an automated physician decision-support tool to predict when adaptive radiation therapy (ART) is required based on dose accumulation from weekly CBCTs.
Methods: A retrospective study of twenty-four stage III lung cancer patients treated to total dose of 60Gy in 30 fractions was identified, of which eleven underwent ART during the course of treatment. For each patient, five to six weekly CBCTs were extracted resulting in a total of 139 training datasets. For the ART cohort, each CBCT was classified as requiring/not requiring ART by comparing the acquisition fraction to the fraction at which the plan was adapted. To quantify the accumulated dose, a synthetic CT was created by deformably registering the weekly CBCT to the original CT, averaging the Hounsfield units, and recalculating the delivered plan. The recalculated dose was deformed back to the original CT dataset to obtain the estimated dosimetric changes caused by the anatomical differences present on each CBCT. These dosimetric changes were quantified into 16 dosimetric features which were used to train a machine learning model based primarily on quadratic discriminant analysis. The model was tested with a 50-fold cross validation to predict whether a replan would have been ordered based on a physicianâ€™s review of the CBCTs.
Results: The testing accuracy of the created model is 89.2%. The area under the ROC curve is 0.9. The sensitivity and specificity are 58% and 98% respectively.
Conclusion: A model capable of predicting a physicianâ€™s decision to replan lung cases based on weekly CBCT imaging has been developed. While this model is based on changes to dosimetric features, the steps to extract these features are all ultimately automatable and can be implemented without negatively impacting the efficiency of the clinical workflow.
Funding Support, Disclosures, and Conflict of Interest: The work was funded in part by a Varian Master Research Agreement.