Deep-learning based localization of Myocardial Fibrosis in CMR images

Localization and quantification of myocardial fibrosis are considered as objective prognostic criteria of successful myocardial revascularization. Visual interpretation of regional left ventricular (LV) abnormalities in clinical practice is typically carried out manually, representing a labor-intensive and time-consuming process that lacks precise quantitative assessment. In this study we have developed a deep learning based semi-automatic tool for myocardium fibrosis localization in 17-segment model in LGE CMRI. The trained CNN cascade model shows a high performance of myocardium and fibrosis segmentation: median Dice similarity coefficient (DSC) 0.84 and 0.74 for myocardium and fibrosis respectively. Testing of the algorithm for localizing and detecting the presence of fibrosis showed an excellent accuracy (F1 0.94 and 0.91 for manual and predicted masks respectively). The high accuracy of fibrosis detection in the 17-segment model confirms that a fibrosis segmentation accuracy (median DSC 0.74) is an acceptable result for achieving high efficiency of the developed tool overall performance.