(Sunday, 7/12/2020) [Eastern Time (GMT-4)]
Room: AAPM ePoster Library
Purpose: We propose a method to calculate the uncertainty in volumetric dose prediction for deep learning systems, with the goal to improve safety in the clinical implementation of artificial intelligence frameworks.
Methods: By utilizing Dropout during inference, where the model’s weights are sampled from a Bernoulli distribution via variational inference, we can approximate a Bayesian method known as a Gaussian process. We apply this method on a U-net style deep neural network architecture for Pareto optimal dose prediction. The dropout rate was set to 0.125, and the model was trained using the Adam optimizer and a learning rate of 1x10?³. In total we used 70 patients, split into 54 training, 6 validation, and 10 testing patients (96 x 96 x 24 dimension arrays at 5 mm³ voxel size). Due to the nature of Dropout and this relation to the variance of prediction (lower dropout rate = lower variance in prediction), we then define a scaling factor, m, such that m*uncertainty is greater than 95% of the error in the validation data.
Results: We found the scaling factor to be m=9.585. On the test data, 97.49% of the error is below m*uncertainty. Using prediction±m*uncertainty as upper/lower bounds, we find that the ground truth dose and resulting dose-volume histograms to match within our uncertainty bounds on our test data. In addition, all predictions have an average mean and max dose error less than 2.8% and 4.2%, respectively, on the PTV and organs at risk, when compared to the test data.
Conclusion: We show a method that can be used so that the uncertainty of deep learning models can be obtained. We characterize a curve to relate the uncertainty to the error in the model. This work can be used to greatly improve the safety of model implementation in a clinical setting.
Funding Support, Disclosures, and Conflict of Interest: National Institutes of Health (NIH) R01CA237269
Keywords
Intensity Modulation, Treatment Planning, Dose
Taxonomy
Contact Email
