Single energy CT-based mass density and relative stopping power estimation for proton therapy using deep learning...

Area of Science:

• Medical Physics
• Radiotherapy
• Artificial Intelligence in Medicine

Background:

• Proton therapy requires accurate patient-specific maps of mass density and relative stopping power (RSP) for dose calculation.
• Current methods using single-energy computed tomography (SECT) have dose calculation uncertainties of 2.5%-3.5% plus a 1 mm margin.
• Enhancing proton dose calculation accuracy with deep learning (DL) on SECT data is a significant area of research.

Purpose of the Study:

• To develop and evaluate a deep learning (DL) method for generating accurate mass density and relative stopping power (RSP) maps from clinical single-energy CT (SECT) data.
• To improve the precision of proton dose calculations in proton therapy treatment planning.

Main Methods:

• Utilized artificial neural networks (ANN), fully convolutional neural networks (FCNN), and residual neural networks (ResNet) to learn the relationship between SECT CT number (HU) and material properties.
• Employed a stoichiometric calibration method and a dual-energy CT (DECT) empirical model as reference methods for performance evaluation.
• Trained DL models using SECT images of a CIRS 062M electron density phantom and tested on CIRS anthropomorphic M701 and M702 phantoms.

Main Results:

• The FCNN model achieved significantly lower mean absolute percentage errors (MAPE) for mass density and RSP maps compared to the SECT stoichiometric method on the M701 phantom.
• For mass density, FCNN achieved MAPEs ranging from 0.39% to 1.57%, while the SECT method yielded MAPEs from 0.99% to 12.96%.
• For RSP maps, FCNN achieved MAPEs ranging from 0.75% to 2.32%, outperforming the SECT reference model's MAPEs of 0.95% to 8.62%.

Conclusions:

• Deep learning neural networks demonstrate strong potential for generating accurate voxel-specific material property information essential for proton therapy.
• The proposed DL-based frameworks offer improved accuracy in estimating mass density and RSP from SECT data compared to conventional approaches.
• This advancement can lead to more precise proton dose calculations and enhanced treatment planning in proton therapy.