Knowledge-based automated planning with three-dimensional generative adversarial networks.

Area of Science:

• Medical Physics
• Radiotherapy
• Artificial Intelligence

Background:

• Automated treatment planning in radiotherapy aims to improve efficiency and plan quality.
• Knowledge-based planning (KBP) leverages past treatment data but often requires manual feature engineering.
• Deep learning offers potential for automated dose prediction without feature engineering.

Purpose of the Study:

• To develop a knowledge-based automated planning (KBAP) pipeline using deep neural networks for predicting 3D dose without feature engineering.
• To evaluate the performance of a novel 3D generative adversarial network (GAN)-based KBP approach within the KBAP pipeline.
• To compare the KBAP pipeline's effectiveness against benchmark deep learning methods and existing clinical plans.

Main Methods:

• Developed a KBAP pipeline integrating a 3D GAN for dose prediction and optimization models for plan generation.
• The 3D GAN predicts dose for the entire 3D CT image, capturing inter-slice correlations, unlike a 2D GAN predicting slice-by-slice.
• Investigated the impact of multiplicative scaling of predicted doses before optimization on clinical criteria satisfaction.

Main Results:

• KBAP plans using scaled 3D GAN predictions achieved 77% clinical criteria satisfaction, outperforming clinical plans (67%).
• Multiplicative scaling improved clinical criteria satisfaction by 11% compared to non-scaled predictions.
• The 3D GAN-based KBAP plans satisfied criteria more frequently than two benchmark deep learning methods.

Conclusions:

• The first knowledge-based automated planning framework utilizing a 3D GAN for dose prediction was successfully developed.
• This 3D GAN-based KBAP framework demonstrated superior performance in satisfying clinical criteria for oropharyngeal cancer treatment plans.
• The developed framework generates more realistic treatment plans compared to previous state-of-the-art approaches.