Long short-term memory networks for proton dose calculation in highly heterogeneous tissues

Area of Science:

• Medical Physics
• Computational Biology
• Radiotherapy

Background:

• Accurate proton dose calculation is crucial for effective radiotherapy.
• Challenging three-dimensional (3D) anatomies pose significant difficulties for traditional dose calculation algorithms.
• Artificial neural networks (ANNs) offer a potential solution for improving dose calculation accuracy and efficiency.

Purpose of the Study:

• To evaluate the feasibility and accuracy of ANNs for proton dose calculations.
• To assess the performance of a novel long short-term memory (LSTM) network in complex 3D treatment scenarios.

Main Methods:

• A novel LSTM network was developed to process 3D CT data slice-by-slice for dose distribution.
• The LSTM approach was compared against Monte Carlo (MC) simulations and pencil beam calculations.
• Studies included artificial phantoms and realistic lung cancer patient cases across various proton beam energies.

Main Results:

• The LSTM model achieved high accuracy, with γ-index pass rates up to 98.57% compared to MC simulations for phantom cases.
• For lung patient cases, pass rates ranged from 94.51% to 98.56% across different energies.
• Independent validation on unseen patient data yielded an average γ-index pass rate of 97.85%.

Conclusions:

• LSTM networks demonstrate suitability for proton dose calculation in radiotherapy.
• Further investigation into model generalization and computational efficiency is recommended for clinical implementation.