Single patient learning for adaptive radiotherapy dose prediction.

Area of Science:

• Medical Physics
• Radiotherapy
• Artificial Intelligence in Medicine

Background:

• Maintaining radiation therapy (RT) plan accuracy is challenging due to anatomical changes during treatment.
• Online adaptation of RT plans is crucial but often manual and time-consuming.
• Deep learning (DL) shows promise for streamlining adaptive radiation therapy (ART) but typically requires large datasets.

Purpose of the Study:

• Introduce a minimalist, patient-specific approach for adaptive dose prediction in ART.
• Train DL models from scratch using only a single patient's initial treatment data.
• Hypothesize that patient-specific DL models will outperform population-based models in dose prediction accuracy.

Main Methods:

• Trained an adaptive population-based (AP) model using data from 33 patients.
• Trained patient-specific (PS) models using initial RT data from 10 individual patients.
• Evaluated model performance on a test set of 26 ART plans using Mean Absolute Percent Error (MAPE) and statistical significance tests.

Main Results:

• PS models achieved a lower average MAPE (4.069%) compared to the AP model (5.759%).
• The difference in MAPE between AP and PS models was statistically significant (p < 0.001).
• Significant differences in mean dose were observed for 12 out of 24 segmented structures.

Conclusions:

• Patient-specific DL models show significant potential for ART, streamlining the training process with minimal data requirements.
• This approach enhances clinical accessibility by enabling personalized dose predictions.
• The single-patient learning strategy offers a promising avenue for personalized cancer treatment in ART.