Improvement of Electrocardiographic Imaging Reconstructions: A Physics-guided AI Approach and an Efficient Method for Training Data Reduction

Date

2023-8-24

Author

Uğurlu, Kutay

Metadata

Show full item record

This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International License.

Item Usage Stats

167
views

19
downloads

Cite This

Non-invasive Electrocardiographic imaging (ECGI) holds promise as a tool that employs distant measurements of body surface potential (BSP) to reconstruct potential distributions on the heart's surface. Due to the nature of the human thorax, the cardiac potentials get smoothed and attenuated. Thus, estimating the epicardial potentials from the BSPs is an ill-posed problem. Hence, regularization is needed. Novel regularization techniques require training data to estimate prior distribution and the common approach is using the whole dataset. The first study in this thesis shows that it is possible to achieve comparable performance to that of all dataset by carefully selecting a small subset of data in the Bayesian Maximum A Posteriori (MAP) solution of the inverse problem. The study proposes two methods on beat selection order and training set expansion termination. The point where the condition number of the covariance stops improving can represent the whole training set’s performance. 26.9% of the dataset resulted in significantly similar metric distributions (p=0.59). The study showed that condition number provides insight about the sufficiency of the training data. The second study uses neural networks (NNs) to learn the implicit prior by solving the problem iteratively with analytical solutions and NN-based denoiser. The method employs decoupled spatiotemporal NN blocks. This outperforms MAP, resulting in more consistent performance. The localization error of the 17 test beats in median (IQR) representation was 14.70(16.60) mm for the zero-order-Tikhonov, 17.00(10.54) mm for the MAP using the same data, and 5.80(8.60) mm for the proposed method, resulting in 38.5% improvement.

Subject Keywords

Electrocardiographic imaging, Deep learning, Neural networks, Training data, Bayesian MAP

URI

https://hdl.handle.net/11511/105244

Collections

Graduate School of Natural and Applied Sciences, Thesis