Deep Learning-Assisted Whole-Body Voxel-Based Internal Dosimetry

Abstract

We propose a novel methodology to conduct whole-body organ-level dosimetry taking into account the heterogeneity of activity distribution as well as patient-specific anatomy using Monte Carlo (MC) simulations and machine learning algorithms. We extended the core idea of the voxel-scale MIRD approach that utilizes a single S-value kernel for internal dosimetry by generating specific S-value kernels corresponding to patient-specific anatomy. In this context, we employed deep learning algorithms to predict the deposited energy distribution, representing the S-value kernel. The training dataset consists of density maps obtained from CT images along with the ground-truth dose distribution obtained from MC simulations. Accordingly, whole-body dose maps are constructed through convolving specific S-values with the activity map. The Deep Neural Network (DNN) predicted dose map was compared with the reference (Monte Carlo-based) and two MIRD-based methods, including single-voxel S-value (SSV) and multiple voxel S-value (MSV) approaches. The Mean Relative Absolute Errors (MRAE) of the estimated absorbed dose between DNN, MSV, and SSV against reference MC simulations were 2.6%, 3%, and 49%, respectively. MRAEs of 23.5%, 5.1%, and 21.8% were obtained between the proposed method and MSV, SSV, and Olinda dosimetry package in organ-level dosimetry, respectively. The proposed internal dosimetry technique exhibited comparable performance to the direct Monte Carlo approach while overcoming the computational burden limitation of MC simulations.