Segmentation of Left Ventricle in Short-Axis MR Images Based on Fully Convolutional Network and Active Contour Model.

Abstract: Left ventricle (LV) segmentation from cardiac MRI images plays an important role in clinical diagnosis of the LV function. In this study, we proposed a new approach for left ventricle segmentation based on deep neural network and active contour model (ACM). The paper proposed a coarse-to-fine segmentation framework. In the first step of the framework, the fully convolutional network was employed to achieve coarse segmentation of LV from input cardiac MR images. Especially, instead of using cross entropy loss function, we propose to utilize Tversky loss that is known to be suitable for the unbalance data-an issue in medical images, to train the network. The coarse segmentation in the first step is then used to create initial curves for ACM. Finally, active contour model was performed to further optimize the energy functional in order to get fine segmentation of LV. Comparative experiments with other state of the arts on ACDCA and Sunnybrook challenge databases, in terms of Dice coefficient and Jaccard indexes, show the advantages of the proposed approach.     «

Abstract: Left ventricle (LV) segmentation from cardiac MRI images plays an important role in clinical diagnosis of the LV function. In this study, we proposed a new approach for left ventricle segmentation based on deep neural network and active contour model (ACM). The paper proposed a coarse-to-fine segmentation framework. In the first step of the framework, the fully convolutional network was employed to achieve coarse segmentation of LV from input cardiac MR images. Especially, instead of u...     »