Segmentation of multi-structures in cardiac MRI using deep learning

Abstract

The heart is one of the most important organs in our body and many critical diseases are associated with its malfunctioning. To assess the risk for heart diseases, Magnetic Resonance Imaging (MRI) has become the golden standard imaging technique, as it provides to the clinicians stacks of images for analyzing the heart structures, such as the ventricles, and thus to make a diagnosis of the patient’s health. However, the examination of these stacks, often based on the delineation of the heart structures, is a tedious and an error prone task due to inter- and intra-variability in the manual delineations. For this reason, the investigation of fully automated methods to support heart segmentation is paramount. Most of the successful methods proposed to solve this problem are based on deep-learning solutions. Especially, encoder-decoder architectures, such as the U-Net (Ronneberger et al., 2015), have demonstrated to be very effective and robust architectures for medical image segmentation. In this work, we propose to use long-range skip connections on the decoder-part of the architecture to incorporate multi-context information onto the predicted segmentation masks and to improve the generalization of the models (see Figure 1). This new module is named Dense-Decoder module and can be easily added to state-of-the-art encoder-decoder architectures, such as the U-Net, with almost no extra additional parameters allowing the model’s size to remain constant. To evaluate the benefits of our module, we performed experiments on two challenging cardiac segmentation datasets, namely the ACDC (Bernard et al., 2018) and the LVSC (Radau et al., 2009) heart segmentation challenges. Experiments performed on both datasets demonstrate that our method leads to an improvement on both the total Average Dice score and the Ejection Fraction Correlation, when combined with state-of-the-art encoder-decoder architectures.