Knowledge-based automated radiation therapy treatment planning utilizing dose prediction with a 2.5D-U-Net

Area of Science:

• Radiation Oncology
• Medical Physics
• Artificial Intelligence in Healthcare

Background:

• Inverse treatment planning (ITP) in intensity modulated radiation therapy (IMRT) is a complex, iterative manual process.
• Optimizing dose constraints is crucial for achieving patient-specific outcomes in IMRT.
• Automating ITP can significantly improve efficiency and consistency in radiation therapy planning.

Purpose of the Study:

• To develop and evaluate a knowledge-based planning (KBP) solution utilizing deep learning to automate ITP for IMRT.
• To assess the feasibility of KBP in generating high-quality treatment plans for prostate and breast cancer patients.

Main Methods:

• A 2.5D-U-Net deep learning model was trained on 72/66 manual plans (MP) for dose prediction.
• The KBP model derived personalized optimization parameters for ITP in 12 test cases each for prostate and breast (60/50Gy prescription).
• Outcomes were compared using DVH metrics, plan quality metric (PQM), and blinded expert evaluation.

Main Results:

• KBP demonstrated median differences (MD) of ≤0.5Gy for prostate organs at risk (OAR) mean doses and ≤0.4Gy for breast OAR mean doses compared to MP.
• For prostate, MD in V56Gy for rectum/bladder was 0.4cc; for breast, MD in V20Gy for ipsilateral lung was -1.0%.
• Expert ratings indicated KBP plans were not markedly worse than MP plans in any test case, with PQM MDs of 0.2% (prostate) and 2.9% (breast).

Conclusions:

• A KBP solution for ITP in IMRT was successfully developed, capable of automatically generating treatment plans of comparable quality to manual planning.
• This automated approach shows potential to streamline and expedite ITP while maintaining high treatment plan quality.
• KBP offers a promising method for enhancing efficiency and consistency in radiation therapy planning.