Whole-body voxel-based internal dosimetry using deep learning

• DOI: 10.1007/s00259-020-05013-4
• Publication Date: 2020-09-01T20:03:35Z
• Authors: Azadeh Akhavanallaf; Iscaac Shiri; Hossein Arabi; Habib Zaidi
• Keywords: Kernel (algebra); Internal radiation

Abstract Purpose In the era of precision medicine, patient-specific dose calculation using Monte Carlo (MC) simulations is deemed gold standard technique for risk-benefit analysis radiation hazards and correlation with patient outcome. Hence, we propose a novel method to perform whole-body personalized organ-level dosimetry taking into account heterogeneity activity distribution, non-uniformity surrounding medium, anatomy deep learning algorithms. Methods We extended voxel-scale MIRD approach from single S-value kernel specific kernels corresponding construct 3D maps hybrid emission/transmission image sets. this context, employed Deep Neural Network (DNN) predict distribution deposited energy, representing S-values, source in center composed human body geometry. The training dataset consists density obtained CT images reference voxelwise S-values generated simulations. Accordingly, are inferred trained model constructed manner analogous voxel-based formalism, i.e., convolving voxel map. map predicted DNN was compared MC two MIRD-based methods, including Single Multiple S-Values (SSV MSV) Olinda/EXM software package. Results exhibited good agreement MC-based serving as mean relative absolute error (MRAE) 4.5 ± 1.8 (%). Bland Altman showed lowest bias (2.6%) smallest variance (CI: − 6.6, + 1.3) DNN. MRAE estimated absorbed between DNN, MSV, SSV respect simulation were 2.6%, 3%, 49%, respectively. dosimetry, proposed SSV, 5.1%, 21.8%, 23.5%, Conclusion DNN-based WB internal comparable performance direct while overcoming limitations conventional techniques nuclear medicine.