3349 - Uncertainty Estimation in Radiotherapy Dose Prediction with Deep Learning for Head and Neck Cancers

L. Chen¹, Z. Wang¹, T. Zhang¹, H. Zhang², X. H. Sun¹, W. Wang¹, J. Duan¹, Y. Gao¹, and L. Zhao¹; ¹Department of Radiation Oncology, Xijing Hospital, Fourth Military Medical University, Xian, Shaanxi, China, ²Ministry of Education Key Laboratory of Intelligent and Network Security, Faculty of Electronics and Information Engineering, Xi’an Jiaotong University, Xian, Shaanxi, China

Purpose/Objective(s): Deep learning predictions are unreliable when the input data is out of the distribution of the training data or corrupted by noise. Detecting such failures automatically is fundamental to integrate deep learning methods into clinical radiotherapy dose prediction. Current approaches for uncertainty estimation of deep learning underestimate the uncertainty due to ignoring prior knowledge about the input data. To address these limitations, we proposed a novel framework for uncertainty estimation based on Bayesian belief networks and Monte-Carlo approximate inference methods. Materials/

Methods: We defined that the total prediction uncertainty consists of the data uncertainty and the model uncertainty. Assumed density filtering (ADF) was used to modify the forward propagation of the deep learning to generate not only output predictions, but also their respective data uncertainties. Bayesian belief networks and Monte-Carlo approximate inference methods were used to represent the model uncertainty. Our framework not only used prior information about the data, e.g., noise in the data, to compute the data uncertainty, but also fully modeled the relationship between data and model uncertainty. We demonstrated both theoretically and experimentally that these two innovations allow our framework to produce higher quality uncertainty estimates than state-of-the-art methods. Our method was trained and evaluated on a public H&N cancer dataset from the OpenKBP 2020 AAPM Challenge.
Results: We showed that our model is capable of generating a reasonable uncertainty map and creating interpretable uncertainties and bounds on the prediction, e.g., mean and total uncertainty. Its prediction accuracy outperforms C3D model, ranking first in OpenKBP 2020 AAPM Challenge, by up to 23% and 15% in dose score and DVH score respectively. For clinical-related DVH dosimetric metrics, the prediction accuracy in terms of mean absolute error (MAE) is 2.04 for D₉₉, 1.54 for D₉₅, 1.87 for D₁, 1.87 for D_mean, 1.89 for D_0.1cc, respectively.
Conclusion: The quantitative results demonstrated that the proposed model achieved more accurate voxel-level dose prediction for head and neck cancers. The model has a great potential to be applied to other disease sites to further improve the quality and efficiency of radiotherapy planning.