Probabilistic 4D predictive model from in-room surrogates using conditional generative networks for image-guided...

Area of Science:

• Medical Imaging
• Radiotherapy Physics
• Computational Biology

Background:

• Respiration-induced organ motion poses challenges for accurate radiotherapy dose delivery.
• Real-time volumetric information is crucial for effective tumor tracking and minimizing damage to surrounding healthy tissues.

Purpose of the Study:

• To develop a novel probabilistic model for volumetric estimation from image-based surrogates during radiotherapy.
• To enable out-of-plane target tracking by predicting organ motion with a scalable predictive horizon.
• To provide uncertainty estimations and allow for subject-specific personalization of the model.

Main Methods:

• A conditional learning framework using 2D surrogate images and pre-operative 3D volumes.
• A sequence-to-sequence temporal mechanism for extrapolating in-time representations from surrogate images.
• Learning population motion fields and associating phase-specific distributions with temporal representations for dense organ deformation recovery.

Main Results:

• The model achieved a mean error of 1.67 ± 1.68 mm for MRI and 2.17 ± 0.82 mm for ultrasound in unseen patient data.
• Personalization of the model resulted in a mean landmark error of 1.4 ± 1.1 mm on volunteer MRI data.
• Demonstrated statistically significant improvements over existing state-of-the-art methods.

Conclusions:

• The proposed probabilistic model effectively estimates organ volumes from image surrogates during radiotherapy.
• The model enables accurate out-of-plane tracking and provides uncertainty quantification.
• Personalization enhances accuracy, offering a robust solution for improving radiotherapy treatments.