A probabilistic deep learning model of inter-fraction anatomical variations in radiotherapy

Area of Science:

• Medical Physics
• Radiotherapy
• Medical Imaging

Background:

• Organ motion during radiotherapy causes dose delivery errors.
• Current motion models are either patient-specific or population-based, with limitations.
• Accurate prediction of inter-fraction variations is crucial for robust treatment planning.

Purpose of the Study:

• To develop a hybrid deep learning approach for predicting patient-specific organ motion.
• To create a daily anatomy model (DAM) that simulates inter-fraction anatomical changes.
• To improve the accuracy and robustness of radiotherapy planning.

Main Methods:

• A deep learning probabilistic framework generating deformation vector fields.
• Utilizing a daily anatomy model (DAM) with few random variables for correlated movements.
• Training the model on 312 CT pairs from 38 prostate cancer patients.

Main Results:

• DAM achieved a DICE score of 0.86 ± 0.05 and prostate contour distance of 1.09 ± 0.93 mm.
• The model outperforms traditional principal component analysis (PCA)-based models.
• Simulated movements accurately reflect the range and frequency of observed daily anatomical changes.

Conclusions:

• The developed DAM accurately predicts patient-specific inter-fraction variations.
• This approach enables robust treatment planning and evaluation against anatomical changes.
• DAM requires only planning CT and organ contours for pre-processing.