Comparison of Deep Learning Models for fast and accurate dose map prediction in Microbeam Radiation Therapy

Area of Science:

• Medical Physics
• Radiotherapy
• Deep Learning

Background:

• Microbeam Radiation Therapy (MRT) utilizes focused synchrotron X-ray microbeams.
• Monte Carlo (MC) simulations are accurate but computationally intensive for MRT dose estimation.
• Deep Learning (DL) offers faster dose prediction in radiotherapy.

Purpose of the Study:

• To compare a Graph-Convolutional-Network (GCN) DL model with a 3D U-Net for MRT dose prediction.
• To evaluate DL model performance against MC simulations in pre-clinical MRT research.
• To analyze dosimetric accuracy and computational efficiency of different DL models.

Main Methods:

• Trained two DL models (GCN and 3D U-Net) using MC-generated 3D dose maps from rat MRT studies.
• Evaluated models against Geant4 MC simulations using Mean Absolute Error, Mean Relative Error, and gamma-index.
• Assessed prediction accuracy in critical regions like tumors, tissue boundaries, and air pockets.
• Compared model execution times and sizes.

Main Results:

• Both DL models achieved comparable overall dosimetric performance.
• Significant differences were observed in accuracy within air pockets.
• Inference times varied, with the 3D U-Net being faster.
• Model choice depends on data structure and time constraints.

Conclusions:

• The GCN model offers flexibility, while the 3D U-Net provides faster execution for MRT dose prediction.
• DL models show promise for efficient dose estimation in pre-clinical MRT.
• Future work should consider model selection based on specific research needs and computational resources.