Knowledge-based versus deep learning based treatment planning for breast radiotherapy.

Area of Science:

• Medical Physics
• Radiation Oncology
• Artificial Intelligence in Medicine

Background:

• Advanced radiotherapy (RT) planning aims to enhance efficiency and quality.
• Knowledge-based planning (KBP) and deep learning (DL) are emerging solutions.
• Direct comparison of KBP and DL models for breast cancer planning is needed.

Purpose of the Study:

• To directly compare KBP and DL models for breast cancer RT planning.
• To evaluate plan quality using the same training, validation, and testing datasets.
• To assess the impact of data cleaning on KBP model performance.

Main Methods:

• Trained two KBP models (clean and non-clean datasets) and one DL U-net model.
• Used 90 RT plans for left-sided breast cancer (15 fractions, 2.6 Gy each) for training/validation.
• Evaluated 15 independent patient plans using dose-volume histogram parameters against clinical plans.

Main Results:

• Both KBP models and the DL U-net model showed small differences in dose calculations compared to clinical plans.
• KBP models underestimated mean heart and lung dose initially but showed higher final mean lung dose.
• The DL U-net model achieved a mean planning target volume dose comparable to clinical plans.

Conclusions:

• KBP and DL models yield clinically acceptable results for breast cancer radiotherapy.
• Data cleaning is not essential for KBP models when a high-quality dataset is used.
• Both KBP and DL offer viable alternatives to manual planning for breast cancer.