LVLNET: Lightweight Left Ventricle Localizer using Encoder-Decoder Neural Network

Abstract

Automatic localization of the left ventricle (LV) is an important preprocessing step in any further analysis or quantification of LV function. Also, LV localization is usually done manually by MRI operator to plan Cardiac Magnetic Resonance Imaging (Cardiac MR) acquisition which can be standardized and automated to reduce the operator's error. In this study, we propose LVLNET; an automatic left ventricle localization approach; which utilizes a lightweight encoder-decoder-like convolutional neural network (CNN). We evaluated our proposed method using three different and independent datasets. The proposed method has estimated the region of interest of the left ventricle with an accuracy of 88% covering more than 90% of the left ventricle voxels. Also, the median distance between the real and estimated centers was 1.12 [0.61–2.38] mm. With the reported results, it is shown that our proposed method had overcome most of the badly annotated images however, considering dynamic movement through series timeframes would boost the resulted accuracy.