Knowledge-based prediction of three-dimensional dose distributions for external beam radiotherapy

Area of Science:

• Medical Physics
• Radiotherapy
• Artificial Intelligence in Medicine

Background:

• Accurate dose prediction is crucial for optimizing radiotherapy plans.
• Knowledge-based treatment planning leverages past clinical data to guide new plans.
• Artificial neural networks (ANNs) offer potential for complex dose prediction tasks.

Purpose of the Study:

• To demonstrate knowledge-based 3D dose prediction for external beam radiotherapy.
• To develop and validate ANNs for predicting dose distributions based on patient-specific parameters.
• To identify potentially suboptimal treatment plans and improve organ-at-risk (OAR) sparing.

Main Methods:

• Trained ANNs using previously treated prostate and stereotactic radiosurgery/radiotherapy (SRS/SRT) plans.
• Input parameters included patient geometry, planning target volume (PTV), and OAR proximity.
• Refined models focused on superior OAR sparing, enabling identification of suboptimal plans for replanning.

Main Results:

• Refined ANNs achieved highly accurate dose distribution predictions for both prostate and SRS cases.
• Prediction bias and interquartile range (IQR) were within acceptable limits across various distances from the PTV.
• Replanning based on model predictions resulted in improved sparing of the rectum and brainstem.

Conclusions:

• Knowledge-based 3D dose prediction using ANNs is highly accurate for radiotherapy.
• This approach can effectively identify and correct suboptimal treatment plans.
• The findings support the integration of AI for enhanced radiotherapy planning and OAR protection.