Learning-Based Stopping Power Mapping on Dual-Energy CT for Proton Radiation Therapy.

Area of Science:

• Medical Physics
• Radiotherapy
• Machine Learning

Background:

• Dual-energy computed tomography (DECT) is used to derive relative stopping power (RSP) maps for proton radiation therapy.
• Physics-based mapping techniques for DECT-derived RSP can be susceptible to image noise and artifacts.

Purpose of the Study:

• To present a noise-robust, learning-based method for predicting RSP maps from DECT data.
• To improve the accuracy and reliability of RSP mapping in proton therapy.

Main Methods:

• A residual attention cycle-consistent generative adversarial network was employed for DECT-to-RSP mapping.
• An inverse RSP-to-DECT mapping was introduced to achieve a 1-to-1 mapping.
• The method was evaluated on 20 head-and-neck cancer patients using a leave-one-out cross-validation strategy.

Main Results:

• The learning-based method achieved an average normalized mean square error of 2.83% and mean error <3%.
• The method maintained performance with added noise, unlike physics-based methods.
• Dose volume histogram metrics showed minimal differences (<0.2 Gy for clinical targets, ~1 Gy for organs at risk).

Conclusions:

• The machine-learning-based method demonstrates high accuracy in predicting RSP maps.
• The approach shows significant potential for improving proton treatment planning and dose calculation.