Hybrid-supervised deep learning for domain transfer 3D protoacoustic image reconstruction

Area of Science:

• Medical Physics
• Artificial Intelligence in Healthcare
• Proton Therapy

Background:

• Protoacoustic imaging offers real-time 3D proton therapy dose verification but suffers from artifacts due to limited acquisition angles, impacting accuracy.
• Addressing the limited view issue is crucial for reliable proton dose verification.

Purpose of the Study:

• To develop a hybrid-supervised deep learning method to overcome the limited view problem in protoacoustic imaging for proton therapy.
• To enhance the accuracy and efficiency of 3D proton dose verification.

Main Methods:

• A two-stage deep learning method, Recon-Enhance, was proposed, utilizing a transformer network for initial reconstruction and a 3D U-net for image enhancement.
• The reconstruction network employed a hybrid-supervised approach with iterative reconstruction, transfer learning, and self-supervision.
• The enhancement network was supervised by ground truth pressure maps.

Main Results:

• The method achieved high accuracy in reconstructing initial pressure maps, with an average RMSE of 0.0292 and SSIM of 0.9618.
• Dose verification demonstrated excellent agreement with ground truth, showing an average RMSE of 0.018 and SSIM of 0.9891.
• Processing time was reduced to 6 seconds, enabling online 3D dose verification.

Conclusions:

• The developed hybrid-supervised deep learning method effectively addresses the limited view issue in protoacoustic imaging.
• This approach significantly improves the accuracy and efficiency of in vivo 3D proton dose verification, enhancing proton therapy precision.
• Protoacoustic imaging shows great promise as a tool for minimizing range uncertainties in proton therapy.