DiffuseRT: predicting likely anatomical deformations of patients undergoing radiotherapy

Area of Science:

• Medical Imaging
• Radiotherapy Physics
• Artificial Intelligence in Medicine

Background:

• Radiotherapy treatment planning requires accurate prediction of anatomical changes.
• Head and neck cancer patients undergoing radiotherapy experience significant anatomical variations.
• Current methods for predicting these changes have limitations.

Purpose of the Study:

• To develop and evaluate deep-learning models for predicting patient-specific anatomical changes during head and neck radiotherapy.
• To assess the performance of different denoising diffusion probabilistic model (DDPM) architectures (image, deformable vector field, hybrid) in generating realistic anatomical variations.
• To investigate the utility of predicted anatomical changes for robust radiotherapy plan optimization.

Main Methods:

• Development of three DDPMs: image model, deformable vector field (DVF) model, and hybrid model.
• Training models on longitudinal cone-beam CT (CBCT) data from head and neck cancer patients.
• Evaluation of model performance using metrics such as Wasserstein distance and comparison with ground truth anatomical changes.
• Application of generated images for robust optimization of proton therapy plans.

Main Results:

• DDPMs successfully predicted anatomical changes, including weight loss in later fractions.
• The image and hybrid models demonstrated better agreement with ground truth anatomical changes than the DVF model (lower Wasserstein distance).
• Utilizing generated images for robust optimization improved the worst-case clinical target volume V95 by 7% compared to standard set-up robustness.

Conclusions:

• The developed DDPMs can generate distributions of anatomical changes similar to those observed in actual patients.
• These models hold significant potential for enhancing robust anatomical optimization in radiotherapy planning.
• Further research can explore variability prediction and clinical integration of these deep-learning approaches.