Using a Residual Neural Network to Predict Radiation-Induced Pulmonary Function Damage
E Wallat1*, M Flakus1, A Wuschner1, J Reinhardt2, G Christensen2, J Bayouth1, (1) University of Wisconsin, Madison, WI, (2) University of Iowa, Iowa City, IA
Presentations
SU-F-BRC-2 (Sunday, 7/10/2022) 2:00 PM - 3:00 PM [Eastern Time (GMT-4)]
Ballroom C
Purpose: To develop a machine learning-based dose response model that predicts pulmonary function following radiation therapy (RT) in low dose regions.
Methods: A residual neural network (RNN) was developed and trained with 212 patient datasets (80/20 train/test split) from a prospective randomized clinical trial approved by our institution’s IRB to predict post-RT pulmonary function. The RNN inputs were the pre-RT function map, dose distribution, and maximum inhale and exhale CT image volumes. Data was normalized using min-max normalization and online augmentation (rotation) was performed during each training batch. Stochastic gradient descent with momentum optimizer was used with an initial learning rate of 0.01; the loss was an asymmetrical structural similarity index measure (SSIM) function designed to increase penalization of under-prediction of functional damage. Dice similarity coefficient (DSC), SSIM, accuracy (ACC), true positive rate (TPR), and true negative rate (TNR) were used as evaluation metrics. RNN performance was evaluated against a previously developed polynomial regression model in low dose regions (<5 Gy) and the entire lung using paired t-tests for comparison. The validation dataset consisted of an additional 25 patient datasets.
Results: In the low dose region, the RNN improved over the polynomial model in TPR, 0.12 to 0.56, and DSC, 0.21 to 0.50, but worsened in TNR, 0.99 to 0.83, ACC, 0.82 to 0.78, and SSIM, 0.90 to 0.89. For the entire lung, the RNN improved over the polynomial model in TPR, 0.12 to 0.63, and DSC, 0.20 to 0.49, but worsened in TNR, 0.98 to 0.78, ACC, 0.82 to 0.75, and SSIM, 0.91 to 0.90. All metrics were significantly (p<0.05) different between models.
Conclusion: The proposed RNN model demonstrated significant improvement in TPR and DSC for both dose regions, which can improve the utility of functional avoidance RT by avoiding the specific regions predicted to decline in function.
Funding Support, Disclosures, and Conflict of Interest: This work was supported by the National Institutes of Health grant CA166703. Joseph Reinhardt is a shareholder in VIDA Diagnostics, Inc. Gary Christensen receives licensing fees from VIDA Diagnostics, Inc. John Bayouth has ownership interest in MR Guidance, LLC, which has business activity with technology used in this study(ViewRay,Inc.).
Keywords
Radiation Therapy, Ventilation/perfusion, Modeling
Taxonomy
TH- Response Assessment: Modeling: Machine Learning
Contact Email