Uncertainty-Aware Deep Learning Reconstruction for 3D Scintillation Dosimetry

Abstract

Scintillation dosimetry enables three-dimensional (3D) dose measurements for quality assurance in radiation therapy through the tomographic reconstruction of scintillation light. With their very fast reconstruction time and high representational capacity, deep learning methods are a promising approach for this reconstruction problem in a limited-view setting where real-time capacity is crucial. However, these reconstruction techniques are hard to interpret and generally give no confidence level on their predictions. In this work, we propose a Bayesian neural network framework for uncertainty estimation in the reconstruction of 3D dose distribution. Results show that predicted uncertainty correlates with reconstruction error, with Pearson correlation coefficients of 0.61 ± 0.06 for the gamma index and 0.94 ± 0.01 for absolute error. It was also found that no statistically significant difference is present between the Bayesian model and its deterministic counterpart, yielding promising results for safer quality assurance.