Convolutional neural network based proton stopping-power-ratio estimation with dual-energy CT: a feasibility study

Area of Science:

• Medical Physics
• Radiotherapy Physics
• Computational Imaging

Background:

• Dual-energy computed tomography (DECT) shows promise for reducing range uncertainties in proton therapy.
• Analytical stopping-power-ratio (SPR) estimation methods struggle with beam-hardening artifacts and CT number variations.
• Accurate SPR is crucial for leveraging the Bragg peak in proton therapy.

Purpose of the Study:

• To develop and evaluate a convolutional neural network (CNN)-based framework for estimating proton SPR.
• To address limitations of analytical methods by accounting for patient geometry and CT number variations.
• To improve the accuracy of SPR estimation in proton therapy.

Main Methods:

• A CNN framework (U-net) was developed to estimate proton SPR.
• The framework was trained using two scenarios: ideal (patient CT images) and realistic (computational phantoms).
• Simulated DECT images and ray-tracing were used to generate ground-truth SPR for evaluation on prostate and head-and-neck patient datasets.

Main Results:

• The CNN framework significantly reduced SPR estimation uncertainty compared to conventional methods.
• Training with computational phantoms (realistic scenario) reduced prostate SPR uncertainty from 1.10% to 0.71% and HN from 2.11% to 1.20%.
• Training with patient images (ideal scenario) yielded even lower uncertainties (0.32% for prostate, 0.42% for HN).

Conclusions:

• CNN-based SPR estimation shows great potential for enhancing accuracy in proton therapy.
• The proposed framework effectively incorporates individual patient geometry information.
• This approach can lead to more precise proton range prediction and improved treatment outcomes.