Improving the generalization of deep learning methods to segment the left ventricle in short axis MR images

Abstract

Cardiovascular disease is one of the major health problems worldwide. In clinical practice, cardiac magnetic resonance imaging (CMR) is considered the gold-standard imaging modality for the evaluation of the function and structure of the left ventricle (LV). More recently, deep learning methods have been used to segment LV with impressive results. On the other hand, this kind of approach is prone to overfit the training data, and it does not generalize well between different data acquisition centers, thus creating constraints to the use in daily routines. In this paper, we explore methods to improve the generalization in the segmentation performed by a convolutional neural network. We applied a U-net based architecture and compared two different pre-processing methods to improve uniformity in the image contrast between five cross-dataset training and testing. Overall, we were able to perform the segmentation of the left ventricle using multiple cross-dataset combinations of train and test, with a mean endocardium dice score of 0.82.