Lightweight and universal deep learning model for fast proton spot dose calculation at arbitrary energies

Area of Science:

• Medical Physics
• Computational Biology
• Radiotherapy Technology

Background:

• Accurate proton spot dose calculation is crucial for advanced radiotherapy techniques like adaptive planning and quality assurance.
• Existing models may lack universality or efficiency for arbitrary proton energies.
• Long short-term memory (LSTM) networks offer potential for modeling complex dose deposition patterns.

Purpose of the Study:

• To propose a lightweight and universal proton spot dose calculation model for arbitrary energies.
• To integrate dose calculation into rapid adaptive planning and quality assurance processes.
• To evaluate the accuracy and efficiency of the proposed model across different anatomical sites.

Main Methods:

• Development of a multi-energy dose long short-term memory (MED-LSTM) model.
• Training and evaluation of the MED-LSTM model on prostate, nasopharynx, and lung cancer cases.
• Assessment of model performance using gamma passing rates under (1%, 3 mm) and (1%, 1 mm) criteria for both spot dose and plan dose.

Main Results:

• MED-LSTM achieved high average gamma passing rates for spot dose calculations (e.g., 99.93% for prostate under (1%, 3 mm)).
• Prostate and lung models showed strong performance for intensity-modulated proton therapy plan dose (e.g., 99.88% for prostate under (1%, 3 mm)).
• Minor deviations were observed in nasopharynx and lung cases due to tissue heterogeneity, indicating areas for further refinement.

Conclusions:

• The MED-LSTM model demonstrates high accuracy and efficiency for proton spot dose calculation across various energies.
• The model's lightweight nature and broad applicability make it promising for clinical integration in proton therapy.
• Further research may focus on addressing deviations in heterogeneous tissues to enhance universality.