BayesDose: Comprehensive proton dose prediction with model uncertainty using Bayesian LSTMs

作     者：Voss, Luke Neishabouri, Ahmad Ortkamp, Tim Mairani, Andrea Wahl, Niklas 

作者机构：Department of Medical Physics in Radiation Oncology German Cancer Research Center – DKFZ Im Neuenheimer Feld 280 Heidelberg69120 Germany Heidelberg Institute for Radiation Oncology – HIRO Im Neuenheimer Feld 280 Heidelberg69120 Germany Ruprecht Karl University of Heidelberg Institute of Computer Science Heidelberg Germany Clinical Cooperation Unit Radiation Oncology German Cancer Research Center – DKFZ Heidelberg Germany  Steinbuch Centre for Computing Karlsruhe Germany Heidelberg Germany HIDSS4Health – Helmholtz Information and Data Science School for Health Karlsruhe Heidelberg Germany  Im Neuenheimer Feld 450 HeidelbergD-69120 Germany 

出 版 物：《arXiv》 (arXiv)

年 卷 期：2023年

核心收录：

主　　题：Forecasting 

摘      要：Purpose: Fast dose calculation techniques are needed in proton therapy, particularly in light of time restrictions in adaptive workflows. Neural network models show the potential to substitute conventional dose calculation algorithms with fast and accurate dose predictions, while lacking measures to quantify an individual prediction’s quality. We propose to use a Bayesian approach to learn uncertainty of AI-based dose prediction. Methods: Our resulting BayesDose-Framework is based on a previously published deterministic LSTM. Similarly, it is trained and evaluated on Monte Carlo beamlet doses simulated on (1) 2500 water phantoms with slab inserts and (2) 1000 geometries extracted from a lung patient for a single initial energy. The network’s weights are parameterized with 2D Gaussian mixture models, and 100 ensemble predictions are used to quantify mean dose predictions and their standard deviation. Generalizability as well as re-training of the model is evaluated on smaller datasets with two different initial energies as well as five additional patients. Results: The averaged predictions of the BayesDose model performed similarly to its deterministic variant and at least as good the original published long short-term memory (LSTM) model. Predictions of the uncertainty (measured through the sampled predictions’ standard deviation) seemed conservative, particularly for the phantom dataset, however regions of high uncertainty correlated spatially with the largest dose differences between the prediction and Monte Carlo calculated reference. Large average uncertainty within a prediction correlates strongly with dosimetric differences (up to ρ = −0.74). This correlation is reduced when applying the model to patient data with unseen HU value ranges. Runtime overhead could be decreased to 9x of a deterministic prediction for an ensemble size of 100 by parallelizing predictions and presampling network weights. Conclusion: Bayesian models for dose prediction can produce fast