Improving the spatial solution of electrocardiographic imaging: a new regularization parameter choice technique for the tikhonov method

Chamorro Servent, Judit; Dubois, Rémi; Potse, Mark; Coudière, Yves

doi:http://dx.doi.org/10.1007/978-3-319-59448-4_28

Improving the spatial solution of electrocardiographic imaging: a new regularization parameter choice technique for the tikhonov method

Citation

Chamorro-Servent J, Dubois R, Potse M, Coudière Y. Improving the spatial solution of electrocardiographic imaging: a new regularization parameter choice technique for the tikhonov method. In: Pop M, Wright GA, editors. Functional Imaging and Modelling of the Heart 9th International Conference Proceedings, FIMH 2017; 2017 June 11-13; Toronto, Canada. [Cham]: Springer; 2017. p. 289-300.(LNCS; no. 10263). DOI: 10.1007/978-3-319-59448-4_28

Abstract

The electrocardiographic imaging (ECGI) inverse problem is highly ill-posed and regularization is needed to stabilize the problem and to provide a unique solution. When Tikhonov regularization is used, choosing the regularization parameter is a challenging problem. Mathematically, a suitable value for this parameter needs to fulfill the Discrete Picard Condition (DPC). In this study, we propose two new methods to choose the regularization parameter for ECGI with the Tikhonov method: i) a new automatic technique based on the DPC, which we named ADPC, and ii) the U-curve method, introduced in other fields for cases where the well-known L-curve method fails or provides an over-regularized solution, and not tested yet in ECGI. We calculated the Tikhonov solution with the ADPC and U-curve parameters for in-silico data, and we compared them with the solution obtained with other automatic regularization choice methods widely used for the ECGI problem (CRESO and L-curve). ADPC provided a better correlation coefficient of the potentials in time and of the activation time (AT) maps, while less error was present in most of the cases compared to the other methods. Furthermore, we found that for in-silico spiral wave data, the L-curve method over-regularized the solution and the AT maps could not be solved for some of these cases. U-curve and ADPC provided the best solutions in these last cases.