Magnetic resonance imaging-based proton dose calculation for pelvic tumors using deep learning.

Area of Science:

• Medical Physics
• Radiotherapy
• Artificial Intelligence in Medicine

Background:

• Magnetic resonance imaging (MRI)-only proton therapy offers superior soft tissue contrast and precise dose delivery.
• Conventional dose calculation is hindered by the lack of electron density information in MRI scans.
• Accurate dose calculation is crucial for effective proton therapy planning.

Purpose of the Study:

• To investigate the feasibility of two deep learning (DL)-based MRI-only proton dose calculation pipelines for pelvic treatments.
• To evaluate the robustness of these pipelines against MRI intensity distortions.
• To compare the accuracy of DL-based dose predictions with Monte Carlo simulations.

Main Methods:

• Two DL pipelines were developed: a two-step approach (MRI to synthetic CT, then dose prediction) and a direct approach (dose prediction directly on MRI).
• Monte Carlo simulations were used to generate ground truth dose distributions for training and validation using MRI-CT data from 120 pelvic patients.
• Pipeline performance was assessed using gamma pass rates and average relative error (ARE) for individual pencil beams and treatment plans, with introduced MRI intensity distortions.

Main Results:

• Both pipelines achieved high gamma pass rates (>99.2%).
• The two-step pipeline demonstrated low ARE (0.11% for pencil beams, 2.63% for treatment plans).
• The direct pipeline showed higher ARE (0.16% for pencil beams, up to 6.11% for treatment plans) but exhibited robustness to MRI intensity distortions.

Conclusions:

• Deep learning-based MRI-only proton dose calculation is feasible for pelvic region treatments.
• The direct DL pipeline shows promise for learning MRI-to-dose mappings, though further optimization is needed.
• The two-step DL pipeline provides accurate proton dose predictions with minimal errors, suitable for clinical implementation.