Real-Time x-Ray Guided Tracking of Multiple Cardiac Structures for Stereotactic Arrhythmia Radiotherapy: A Feasibility Study
Abstract
Purpose
Stereotactic Arrhythmia Radioablation(STAR) is a non-invasive method to treat cardiac arrhythmias by targeting aberrant cardiac tissue, however accurate dose delivery is challenging, given the proximity of the target to vital cardiac structures and the complex cardiorespiratory motion, making real-time tracking desirable. This work describes the feasibility of AI-powered real-time x-ray guided multi-structure tracking to improve STAR.
Methods
X-ray guided tracking was enabled using a conditional-Generative Adversarial Network(cGAN). The cGAN was selected based on demonstrated efficacy in markerless tracking for other sites. Six different anatomies were created using the 4D-XCAT digital phantom, with cardiorespiratory motion traces derived from the PhysioNet CEBS database. Training data was created from 1-minute CEBS patient traces which were used to create 10 volumes of a 4D-CT, and corresponding heart, left-ventricle(LV) and clinical target volume(CTV) segmentations. The 3D heart, LV and CTV segmentations were forward projected to create ground truth 2D segmentations. Training data for anatomy-specific cGANs were generated by deforming and forward projecting each 3D-CT volume from the 4D-CT volumes to create 7200 unique Digitally Reconstructed Radiographs(DRRs) per epoch(0.5° angular resolution). Testing data was generated using non-overlapping 5-minute traces to create 1050 XCAT volumes, with each volume being forward projected at unique projection angle simulating kV image acquisition. The cardiac tracking accuracy was measured by comparing the cGAN output to the ground truth cardiac segmentations using mean surface distance(MSD) and centroid error. The cGAN accuracy was compared to “no-tracking” segmentations from mid-ventilation respiration phase.
Results
cGAN tracking MSD for the heart(2.1±0.8mm), LV(2.8±1.2mm), and CTV(3.9±3.0mm), were lower than the no-tracking MSD heart(5.0±2.3mm), LV(6.0±2.3mm), and CTV(5.2±2.3mm). Similarly, cGAN tracking centroid errors for the heart(2.5±1.5mm), LV(3.6±2.1mm), and CTV(5.6±3.9mm) were lower than no-tracking for heart(5.2±3.0mm), LV(6.1±2.9mm), and CTV(7.5±3.5mm).
Conclusion
Feasibility of real-time cardiac tracking of multiple cardiac structures has been demonstrated, enabling improved STAR treatments.