New paper on Computer Methods and Programs in Biomedicine

Late Gadolinium Enhancement (LGE) is a standard technique for Cardiovascular Magnetic Resonance (CMR), allowing the non-invasive analysis of myocardial issues. The main issue of classic LGE is that blood can appear as bright as scar regions, which complicates the interpretation of results. In the last decade, multiple “Dark-Blood” LGE (DB-LGE) approaches have been proposed to tackle this issue and increase contrast in CMR. Besides the improvement of diagnosis, DB-LGE also offers a new opportunity to train deep learning models, most notably Convolutional Neural Networks (CNN).

In this work, we developed a complex system composed of two U-Net CNNs in cascade: the first one is used to segment the myocardium, and the second one segments the scar highlighted by DB-LGE. By using this approach, we can precisely assess the extent of myocardial infarction and transmurality and automatically generate a detailed report. Our method can give a response in a very short time, and thanks to U-Net’s semantic segmentation, the outcome is easy to interpret.

This research spanned over three years and represents the result of a large cooperation between Ca’ Foscari, the Department of Cardiology of the Istituto Auxologico Italiano, the University of Milano-Bicocca, the University College of London, and the American National Institute of Health. The paper was published on Elsevier’s Computer Methods and Programs in Biomedicine.