Predictive online 3D target tracking with population-based generative networks for image-guided radiotherapy

Area of Science:

• Medical Imaging
• Radiotherapy Physics
• Machine Learning in Medicine

Background:

• Respiratory motion significantly challenges accurate tumor targeting in image-guided radiotherapy.
• Real-time volumetric data is crucial for improving target tracking, enhancing treatment efficiency, and minimizing healthy tissue toxicity.

Purpose of the Study:

• To develop a novel population-based generative network for 3D target location prediction from 2D image surrogates.
• To enable real-time, out-of-plane tracking of treatment targets during free-breathing radiotherapy.

Main Methods:

• A generative network was trained to represent 3D non-rigid deformations and predict target motion ahead of time.
• The model utilizes a baseline patient anatomy volume for error correction, mitigating system latency.
• The approach does not require supervised data like ground-truth registration fields or organ segmentation.

Main Results:

• The model achieved mean landmark errors of 1.8 mm (volunteer MRI), 2.4 mm (patient MRI), and 5.2 mm (ultrasound).
• Evaluation on diverse datasets (4D MRI, ultrasound) demonstrated robustness across different acquisition protocols.
• The method enables 3D target tracking from single-view slices without prior subject-specific 4D data.

Conclusions:

• The proposed model offers advantages over existing methods, including an explainable latent space with respiratory phase discrimination.
• It demonstrates strong generalization capabilities, eliminating the need for inter-subject correspondences.
• With a rapid 8 ms inference time, the network accurately predicts anatomical changes and tracks tumors in real time, showing significant improvements.