Whole-body voxel-based internal dosimetry using deep learning.

Area of Science:

• Medical Physics
• Radiological Sciences
• Computational Biology

Background:

• Precision medicine relies on accurate patient-specific dose calculations for radiation therapy risk-benefit analysis.
• Monte Carlo (MC) simulations are the gold standard but are computationally intensive.
• Existing dosimetry methods have limitations in handling complex anatomical and distributional heterogeneities.

Purpose of the Study:

• To develop a novel deep learning (DL) algorithm for whole-body, organ-level personalized dosimetry.
• To account for patient-specific anatomy, activity distribution heterogeneity, and surrounding medium non-uniformity.
• To provide a faster and equally accurate alternative to MC simulations for internal dosimetry.

Main Methods:

• Extended the voxel-scale MIRD approach using patient-specific anatomy to create 3D dose maps.
• Employed a Deep Neural Network (DNN) to predict specific S-value kernels from CT-derived density maps.
• Trained the DNN using MC-generated S-values and validated against MC simulations and conventional methods (MSV, SSV, Olinda/EXM).

Main Results:

• The DNN model accurately predicted specific voxel S-value kernels with a mean relative absolute error (MRAE) of 4.5% compared to MC.
• DNN demonstrated the lowest dose bias (2.6%) and smallest variance in Bland and Altman analysis.
• Organ-level dosimetry MRAE for DNN was 5.1%, significantly outperforming MSV (21.8%) and SSV (23.5%).

Conclusions:

• The proposed DNN-based whole-body internal dosimetry method achieves performance comparable to MC simulations.
• This DL approach overcomes the computational limitations of traditional dosimetry techniques in nuclear medicine.
• The method enables more efficient and accurate personalized risk-benefit analysis in precision medicine.