Machine learning-based automatic proton therapy planning: Impact of post-processing and dose-mimicking in plan...

Area of Science:

• Medical Physics
• Radiation Oncology
• Computational Biology

Background:

• Automated treatment planning, including knowledge-based planning (KBP), is increasingly adopted in clinical practice.
• KBP aims to reduce variability, accelerate optimization, and enhance plan quality in radiation therapy.
• Intensity-modulated proton therapy (IMPT) offers precise dose delivery, but plan optimization remains complex.

Purpose of the Study:

• To evaluate the feasibility and quality of IMPT plans generated using four distinct KBP pipelines.
• To assess the impact of post-processing and dose mimicking strategies on IMPT plan robustness and quality.
• To compare KBP-generated IMPT plans against manually generated plans.

Main Methods:

• Sixty oropharyngeal cancer patient datasets were used, with data split into training, validation, and testing sets.
• Four KBP pipelines were developed, integrating dose prediction, post-processing (PP), and dose mimicking (DM) algorithms (RayStation-based mimicking (RSM) and isodose-based mimicking (IBM)).
• Plan quality and robustness were evaluated using dose-volume histogram (DVH) metrics in nominal and worst-case scenarios, comparing KBP plans to manual plans.

Main Results:

• Nominal IMPT plans from KBP pipelines demonstrated comparable target coverage and improved organ-at-risk (OAR) sparing versus manual plans.
• Overly optimistic post-processing in KBP compromised plan robustness, with 65% of patients not meeting robustness criteria.
• No post-processing with RayStation-based mimicking (NPP-RSM) offered the best balance between robustness and OAR sparing.

Conclusions:

• Post-processing and dose mimicking are critical for generating robust and deliverable IMPT plans from machine learning-predicted doses.
• Clinical implementation of KBP requires modification of default parameters and a feedback loop to ensure robustness.
• Optimized KBP strategies can achieve clinically acceptable IMPT plan quality comparable to manual planning.