Automated Predictive Models for Lung ART
J Kavanaugh1*, G Hugo1, Z Ji1, J Fontenot4, (1) Washington University School of Medicine, St. Louis, MO, (2) Mary Bird Perkins Cancer Center, Baton Rouge, LA
Presentations
SU-D-TRACK 4-2 (Sunday, 7/25/2021) 2:00 PM - 3:00 PM [Eastern Time (GMT-4)]
Purpose: Mid-treatment anatomic changes for patients undergoing radiation therapy for lung cancer can alter the accuracy of delivered vs. planned dose. Changes visible on the daily conebeam CT (CBCT) can be classified into five types: tumor regression, tumor alignment, nodal target alignment via two surrogates (carina and thoracic vertebrae) and lung density. To maintain the intended treatment quality, adaptive radiotherapy (ART) modifies the treatment plan to account for each type of change. However, the evaluation on when to adapt is completed manually, resulting in a subjective and inconsistent application. To address these limitations, five predictive ART models were created to identify when to implement ART.
Methods: Models were trained utilizing 1158 CBCT datasets from 43 patients treated 2010–2018. The “ground-truth” need for ART was evaluated on each CBCT by two trained clinical staff, with ART indicated when anatomic deviations exceeded predefined tolerances across three consecutive days. The rate and magnitude of anatomic changes were quantified using a subset of 368 features derived from image registration similarity criteria applied between the simulation CT and on-treatment CBCT. These quantifying criteria, along with the “ground-truth” ART evaluation, were input data for predictive logistic regression models, with one model for each type of anatomic changes. The models were evaluated with a 10-fold cross-validation, focusing on the specificity, sensitivity, and first identified fraction compared to “ground-truth”.
Results: The models identified ART with a range of sensitivity (78.9%–100%) and specificity (93.5%-100%) across five types of anatomic changes. Model accuracy to predict the specific fraction for ART ranged from -3.58±5.35 to 0.14±7.54 fractions, with negative numbers indicating ART prediction occurring before the “ground-truth”.
Conclusion: A method for automated ART predictive modeling in lung cancer patients was developed and provided a high degree of accuracy to predict the need for ART for each individual fraction.
Funding Support, Disclosures, and Conflict of Interest: Washington University receives research support from Siemens, Varian Medical Systems, and ViewRay. Dr. Kavanaugh reports personal fees from Varian Medical Systems outside the scope of this work. Dr. Hugo reports personal fees from Varian Medical Systems outside the scope of the present work.
Handouts
Keywords
Lung, Radiation Therapy, Modeling
Taxonomy
IM/TH- Cone Beam CT: Machine learning, computer vision
Contact Email