Fully automatic segmentation of post-infarction fibrosis in post-contrast magnetic resonance images of heart : preliminary study

Volume of myocardial fibrous tissue have been proposed as an objective prognostic criteria of a successful myocardial revascularization. Manual delineation of fibrosis tissue in magnetic resonance images is a very tedious task even when a contrast enhancement is used. Convolutional neural networks have been widely applied for hard semantic segmentation tasks. Previously, a deep learning-based approach has been successfully developed for segmentation of left ventricle myocardium fibrosis in three dimensional magnetic resonance images with late gadolinium enhancement. However, for the widely used two dimensional images with late gadolinium enhancement, there have not been proposed a fully automated segmentation tool yet. In the present preliminary study, we intended to design a fully automated instrument with the aim of segmenting multiple regions in 2D LGE MR images of heart i.e. left ventricle, myocardium and fibrotic areas in myocardium. The results demonstrate that fully automatic segmentation of heart structures, including myocardial fibrosis, is feasible in 2D LGE MR images with a promising accuracy. Segmentation accuracy and homogeneity may benefit from the further addressing a class unbalance issue.