High quality proton portal imaging using deep learning for proton radiation therapy: a phantom study

Area of Science:

• Medical Physics
• Radiotherapy Technology
• Artificial Intelligence in Medicine

Background:

• Proton portal imaging (PPI) provides valuable data for validating tumor location during proton radiotherapy.
• Current PPI techniques suffer from poor contrast and spatial resolution, limiting their clinical utility.
• kV digitally reconstructed radiographs (DRRs) offer higher quality but are not directly acquired during treatment.

Purpose of the Study:

• To develop and evaluate a deep-learning-based method for enhancing the quality of proton portal images (PPIs).
• To improve the spatial resolution and contrast of PPIs using kV digitally reconstructed radiographs (DRRs) as a reference.
• To assess the feasibility of using deep learning for real-time, in-treatment imaging in proton radiotherapy.

Main Methods:

• A residual generative adversarial network (GAN) framework was employed to learn the mapping between PPIs and DRRs.
• Residual blocks were utilized to focus the network on structural differences between the image types.
• The model was trained on 149 PPI/DRR pairs and validated using 30 test images with a six-fold cross-validation scheme.

Main Results:

• Qualitative assessment showed enhanced spatial resolution and fine detail in corrected PPIs compared to original PPIs.
• Quantitative metrics demonstrated high accuracy, including a normalized mean error (NME) of -0.1% and a structural similarity (SSIM) index of 0.987.
• The method successfully generated high-quality corrected PPIs that closely resemble DRRs.

Conclusions:

• The proposed deep learning approach effectively enhances PPI quality, significantly improving spatial resolution and detail.
• This method demonstrates the potential of deep learning to make PPI a viable tool for patient position verification in proton radiotherapy.
• The enhanced PPI offers a promising alternative to traditional orthogonal X-rays or cone-beam CT for beam's-eye-view imaging during treatment.