A deep learning model for inter-fraction head and neck anatomical changes in proton therapy

Area of Science:

• Medical Imaging and Radiation Oncology
• Artificial Intelligence in Healthcare
• Computational Anatomy

Background:

• Inter-fraction anatomical variations significantly impact radiotherapy accuracy in head and neck cancer patients.
• Accurate prediction of these anatomical changes is crucial for adaptive radiotherapy and robust treatment planning.
• Current methods for predicting anatomical changes often lack the precision required for real-time adaptation.

Purpose of the Study:

• To evaluate a probabilistic deep learning algorithm, the daily anatomy model (DAM_HN), for predicting inter-fraction anatomical variations.
• To assess the model's performance in reconstructing patient anatomy and generating realistic anatomical changes during treatment.
• To explore the utility of DAM_HN in enhancing radiotherapy planning and robustness evaluation.

Main Methods:

• Developed a daily anatomy model (DAM_HN) using a variational autoencoder architecture to approximate conditional probability distributions.
• Trained the model on 315 planning CT (pCT) and repeat CT (rCT) image pairs from 93 head and neck patients.
• Assessed performance using DICE scores (0.83) and normalized cross-correlation (0.60) on a held-out test set.

Main Results:

• The DAM_HN achieved a DICE score of 0.83 and a normalized cross-correlation of 0.60, indicating high reconstruction accuracy.
• Generated anatomical changes, including volume and center of mass shifts for organs like parotid glands and spinal cord, closely matched observed clinical data.
• The model successfully replicated known anatomical shifts, such as medial parotid gland movement, aligning with literature findings.

Conclusions:

• The probabilistic deep learning algorithm (DAM_HN) effectively generates realistic anatomical variations encountered during head and neck radiotherapy.
• This model shows significant potential for applications in robust optimization and library-based treatment planning.
• DAM_HN offers a valuable tool for improving treatment robustness and adaptability against inter-fractional anatomical changes.