Author: Vicky Chin, Mark Gardner, Nicholas Hindley, Paul J. Keall, Adam Mylonas 👨🔬
Affiliation: Image X Institute, Faculty of Medicine and Health, The University of Sydney 🌍
Purpose: Stereotactic Arrhythmia Radioablation (STAR) is a non-invasive method to treat cardiac arrhythmias by targeting aberrant electrical conduction regions in the heart. Targeting is challenging given the radiosensitivity of the heart and the complex combined cardiorespiratory motion. For most treatments, no real-time tracking is used. This work describes proof-of-principle of an AI-powered real-time x-ray guided tracking to improve STAR.
Methods: The x-ray guided tracking was enabled using a conditional Generative Adversarial Network (cGAN). The cGAN was selected based on its intrinsic applicability to image segmentation problems and demonstrated efficacy for non-cardiac target tracking.
Data was created for eight different anatomies using the 4D extended cardiac-torso (4D-XCAT) digital phantom, with cardiorespiratory motion derived from the Combined measurement of ECG, Breathing and Seismocardiogram (CEBS) database patient traces. The use of the XCAT phantom allows for accurate ground truth heart locations. Training data was created from 1-minute CEBS patient segments for each anatomy which were used to create 4D-CT planning volumes (and corresponding heart segmentations).
Anatomy-specific cGANs were trained by deforming and forward projecting 4D-CT volumes to create 7200 unique Digitally Reconstructed Radiographs (DRRs) per epoch.
Cardiac tracking was tested using 5-minute CEBS patient segments to create 3150 unique volumes which were forward projected to create 3150 DRRs per anatomy. 3D-XCAT heart segmentations were forward projected to create ground truth 2D heart segmentations.
The cardiac tracking performance was measured by comparing the cGAN output to the ground truth heart segmentations using the centroid error, dice and mean surface distance (MSD).
Results: The cardiac-tracking centroid error in the x (Sup-Inf) and y (Left-Right/Ant-Post) directions was (mean±stdev) -0.2±2.6mm and 0.2±1.7mm respectively and 0.94±0.02 and 2.9±1.0 mm for dice and MSD respectively. The submillimeter centroid accuracy demonstrates the proof-of-principle of real-time cardiac-tracking.
Conclusion: We have demonstrated the potential for real-time cardiac tracking for improving STAR treatments.