The importance of evaluating the complete automated knowledge-based planning pipeline

Area of Science:

• Medical Physics
• Computational Biology
• Radiation Oncology

Background:

• Knowledge-based planning (KBP) in radiation therapy involves complex pipelines combining prediction and optimization methods.
• Evaluating the interplay between different prediction algorithms (Generative Adversarial Network - GAN, Random Forest - RF) and optimization strategies (Inverse Planning - IP, Dose Mimicking - DM) is crucial for improving treatment plan quality.

Purpose of the Study:

• To determine how different prediction and optimization method combinations in two-stage KBP pipelines affect radiation therapy treatment plan outcomes.
• To benchmark the performance of four distinct KBP pipelines (GAN-IP, GAN-DM, RF-IP, RF-DM) against clinical plans and evaluate prediction errors.

Main Methods:

• Trained GAN and RF models on 130 treatment plans for dose prediction.
• Applied prediction models to 87 out-of-sample patients to generate predicted dose distributions.
• Integrated predicted doses into IP and DM optimization models to create four KBP pipelines.
• Benchmarked pipeline performance against clinical plans using established criteria and evaluated prediction errors (mean absolute error).

Main Results:

• The GAN-IP pipeline generated plans that met clinical criteria most often (78%), outperforming other KBP pipelines.
• Despite GAN-IP's success, GAN predictions did not consistently lead to the best-performing second-stage optimization outcomes.
• RF-IP and RF-DM plans showed notable improvements over GAN-DM, indicating that prediction method choice impacts optimization effectiveness.
• RF predictions exhibited lower mean absolute error (3.6 Gy) compared to GAN predictions (3.9 Gy).

Conclusions:

• The combination of prediction and optimization methods in KBP pipelines significantly influences radiation therapy treatment plan quality.
• The GAN-IP pipeline demonstrated superior performance, but the overall effectiveness depends on the specific pairing of prediction and optimization algorithms.
• Further research is needed to optimize the synergy between advanced prediction techniques and various optimization strategies for consistent, high-quality treatment planning.