Automated plan generation for prostate radiotherapy patients using deep learning and scripted optimization

Area of Science:

• Medical Physics
• Radiotherapy
• Artificial Intelligence in Medicine

Background:

• Treatment planning in radiotherapy is a complex and time-consuming process.
• Automation of radiotherapy treatment planning can improve efficiency and consistency.

Purpose of the Study:

• To develop and evaluate a "single-click" automated workflow for prostate radiotherapy treatment planning.
• To integrate a deep learning model (DLM) prediction into a commercial treatment planning system (TPS) for autoplanning.

Main Methods:

• A ResUNet deep learning model (DLM) was trained to predict 3D dose distributions using 3D contours.
• Predicted dose distributions were converted into treatment plans using TPS objectives and inverse optimization.
• An automated workflow was scripted and deployed within a commercial TPS.

Main Results:

• Automated plans showed high agreement with clinical plans for planning target volumes (e.g., V100% difference of 0.4%).
• Objectives for bladder and rectum dose agreed within -6.1%.
• The automated plan generation process, including DLM prediction and optimization, took approximately 15 minutes.

Conclusions:

• A fully deployed, single-click automated radiotherapy treatment planning workflow using a DLM was successfully implemented in a commercial TPS.
• The generated autoplans were found to be non-inferior to manually created clinical plans.
• This automated approach offers a promising solution for efficient and accurate radiotherapy treatment planning.