Solving the the inverse problem of electrocardiography requires precise knowledge of ECG electrode positions. Manual segmentation is time-consuming and error-prone. The distinct shape of ECG electrodes and then proximity to the torso surface may allow for an automatic detection.
A template matching approach based on cross-correlating the image with a (rotated) 3D template of an electrode showed the following problems:
-Intensity inhomogeneities in MR images make it difficult to define a global threshold that separates electrodes from background in the correlation output, even when using zero-mean normalized cross-correlation.
-The template still needs to be optimized for maximum gain in SNR.
The project involves the following steps:
-Manually segment many electrodes and derive and optimal optimal template by averaging individual electrode segmentations.
-Find a robust way to normalize the cross-correlation or to define a threshold.
-Evaluate the use of other feature descriptors (for example the histogram of oriented gradients).
-Combine several feature descriptors to crate a stronger classifier.
-Make use of the regular spatial arrangement of electrodes within one electrode strip.