Left atrial segmentation from LGE MRI using a convolutional neural network

Segmentacija dijelova srca iz kasno postkontrastno snimljenih slika magnetne rezonance (LGE MRI) predstavlja važan dio svakodnevne medicinske prakse kako bi se pravilno identificirala fibroza miokarda te drugih bolesti srca i kardiovaskularnog sustava. Međutim, automatska segmentacija i dalje predstavlja veliki izazov zbog heterogene raspodjele intenziteta i nejasnih granica između dijelova srca. U ovom radu predstavljena je automatska metoda segmentacije lijevog atrija iz trodimenzionalnih (3D) LGE MRI slika temeljena na dubokom učenju. Segmentacija lijevog atrija izvedena je pomoću konvolucijske neuronske mreže korištenjem U – Net modela. U-Net model treniran je pomoću izrazito malog skupa podataka od 6 3D LGE MRI slika. Unatoč tome, uspješno je ostvaren dice koeficijent od 0.67. Prema tome, u radu je pokazano da originalna U-Net arhitektura može ostvariti uspješan rezultat za segmentaciju lijevog atrija iz LGE-MRI slika upotrebom malenog broja podataka za učenje mreže.Segmentation of the heart chambers from late gadolinium enhancement magnetic resonance images (LGE MRI) is an important part of everyday medical practice. It allows proper identification of myocardial fibrosis and other various heart and cardiovascular system diseases and conditions. However, automatic segmentation remains a major challenge due to the heterogeneous distribution of intensities and blurred boundaries between parts of the heart. This paper presents an automatic left atrial segmentation method from three-dimensional (3D) LGE MRI images based on deep learning. Left atrial segmentation was performed using a convolutional neural network using the U - Net model. The U-Net model was trained using an extremely small data set of 6 3D LGE MRI images. Despite this, a dice coefficient of 0.67 was successfully achieved. Therefore, the paper shows that the original U-Net architecture can successfully segmentation the left atrium from LGE-MRI images using a small amount of data for network learning