Automatic scar segmentation on late gadolinium enhancement cardiovascular magnetic resonance images of patients with Tetralogy of Fallot

Abstract

Tetralogy of Fallot (TOF) is the most common form of cyanotic congenital heart disease, and one of the first to be successfully repaired by congenital heart surgeons. Even though the long-term outcome is favourable in most cases, several authors have drawn special attention to the unexpected occurrence of arrhythmias and sudden death originated by the alteration of the electrophysiological function of the heart due to fibrosis and the scars produced in the surgical repair. However, currently there is no reliable way to assess which patients are at risk. Taking on this challenge, the research group led by Babu-Narayan S. is trying to relate fibrosis patterns in MRI late gadolinium enhancement patient images with late outcome to determine the prognosis of those patients who underwent TOF surgical repair. To enable the latter, it is crucial to obtain accurate cardiac scar segmentations from the patient images. Current methodology is time consuming, not reliable, often require manual refinement and the turnover time would never be applicable to clinical workflow. This underlines the need for the development of a reliable algorithm capable of removing observer bias and with clinically acceptable accuracy. With that objective, we explored several algorithms for automatic scar segmentation and compared its performance: Gaussian Mixture Models, Gaussian Mixture Models with Full-Width-at-Half-Maximum, Gaussian Mixture Models with GrowCut, 3D automatic GrabCut and n-SD. The quantitative analysis and qualitative analysis of the results based on 13 patient datasets revealed Gaussian Mixture Models with Full-Width-at-half-Maximum as the most promising approach for automatic scar segmentation and discarded the use of the n-SD approach. Nevertheless, further work should be carried out with improved data and larger sample size to draw reliable conclusions.