Background:
Non-invasive electrocardiographic imaging (ECGI) is a technique used to reconstruct the electrical activity of the heart using measurements recorded from body-surface electrodes. A typical workflow requires heart and torso geometry from tomographic data, along with 3D electrode locations. This is not part of the routine arrhythmia workflow, and additional acquisition burdens the patient.
Objective:
Our goal is to evaluate the accuracy of the ECGI technique when using a 3D camera to localize the electrodes, with the goal of reducing the procedure's overhead. We compare against results obtained when using traditional imaging.
Method:
Electrode localization is performed by recording multiple observations of individual electrodes in every frame of the recorded video from the 3D camera. A pose graph is constructed using the observations as nodes and the rigid transformation between observations as the edges. A global optimization is applied over this graph to minimize error of all transforms, and the results are combined into a single observation for each lead, generating the camera-based torso. The camera torso is registered to cardiac scan data via corresponding anatomical landmarks that are detected in the camera data using a convolutional neural network and annotated in the scan data.
Results:
The proposed method was validated on 8 clinical patients. The errors between the camera-based torso and the torso from scan data were measured using the Euclidean distance. When measured across all the patients, the mean distance error between corresponding electrodes was 50mm, which is the measured distance between one lead and its neighbor.
Conclusion:
The camera-based torso was accurate enough to the torso generated from the scan data that the ECGI technique could be performed successfully. Although further work can be done to reduce the error between the two types of torsos, the current result is accurate enough to generate valid activation maps.