Deep-dose: a voxel dose estimation method using deep convolutional neural network for personalized internal dosimetry

Area of Science:

• Medical Physics
• Radiological Sciences
• Artificial Intelligence in Medicine

Background:

• Accurate personalized dosimetry is essential for advancing personalized medicine.
• Direct Monte Carlo simulation is a precise but computationally intensive dosimetry method.
• Existing voxel-based dosimetry techniques face limitations in speed and accuracy.

Purpose of the Study:

• To develop a deep convolutional neural network (CNN) for rapid and accurate voxel dose prediction in personalized dosimetry.
• To evaluate the performance of the CNN against direct Monte Carlo simulations and conventional methods like Voxel S-Value (VSV) kernel convolution.

Main Methods:

• Utilized PET and CT image patches as input for a CNN model.
• Trained the CNN using ground truth data from direct Monte Carlo simulations.
• Compared CNN-predicted dose rate maps with ground truth and VSV kernel convolution results.
• Assessed organ absorbed dose errors in whole-body dosimetry studies.

Main Results:

• The CNN achieved high accuracy, with voxel dose rate errors of 2.54% ± 2.09%, closely matching direct Monte Carlo.
• The VSV kernel approach showed significantly higher voxel errors (9.97% ± 1.79%).
• In whole-body dosimetry, the CNN demonstrated superior performance with average organ absorbed dose errors of 1.07% compared to 9.43% for VSV and 34.22% for OLINDA/EXM.

Conclusions:

• The proposed CNN-based dosimetry method offers a significant improvement over conventional techniques.
• This AI approach provides dosimetry results comparable to direct Monte Carlo simulations.
• The CNN method drastically reduces calculation time, making it a viable alternative for clinical applications.