DeepDose: a robust deep learning-based dose engine for abdominal tumours in a 1.5 T MRI radiotherapy system

Area of Science:

• Medical Physics
• Radiotherapy
• Artificial Intelligence

Background:

• Accurate dose calculation is crucial for effective radiotherapy.
• MRI-guided radiotherapy offers real-time imaging for precise tumor targeting.
• Deep learning presents a promising avenue for improving radiotherapy dose calculations.

Purpose of the Study:

• To develop and validate a deep learning-based framework for accurate dose calculations in abdominal radiotherapy using a 1.5 T MRI system.
• To assess the performance and generalizability of the framework across different abdominal tumor sites.

Main Methods:

• A convolutional neural network (CNN) was trained using the DeepDose framework on dose data from individual multi-leaf-collimator segments.
• Training data comprised 176 patient fractions from prostate, rectal, and oligometastatic abdominal tumors treated on an Elekta MR-linac.
• Ground truth dose distributions were generated using a Monte Carlo dose engine with 1% statistical uncertainty per segment.

Main Results:

• The deep learning framework achieved 99.4% ± 0.6% agreement with the 3%/3 mm gamma index for whole-plan dose prediction on 20 independent test fractions.
• The average dose difference per segment was 0.3% ± 0.7%.
• Excellent agreement was also observed for additional cervical and pancreatic cancer cases (99.9% and 99.8% respectively).

Conclusions:

• The developed deep learning dose engine provides highly accurate dose distributions for abdominal tumors treated on an MR-linac.
• The framework demonstrates robust performance and generalizability across diverse abdominal tumor sites.
• This approach has the potential to significantly enhance dose calculation accuracy and efficiency in MRI-guided radiotherapy.