Multiple Attention Fully Convolutional Network for Automated Ventricle Segmentation in Cardiac Magnetic Resonance Imaging

Automated ventricle segmentation for cardiac cine magnetic resonance imaging (MRI) images is a challenging problem. Variability of the shape of the heart and similar intensity with adjacent tissues are impediments that frustrate automated segmentation. Ventricle segmentation still remains an important step for calculating clinical indices like ejection fraction and ventricular volume. In this study, we propose a multiple attention fully convolutional network (MAFCN) based on attention mechanism for ventricle segmentation. Specifically, we introduce three types of attention modules on top of a modified fully convolutional network (FCN), including spatial attention, channel attention and region attention. The spatial attention module selectively aggregates the features at each position, the channel attention module focuses on integrating associated features among all channels, and the region attention module highlights useful feature regions from the whole feature maps. Meanwhile, these three attention modules are modular and can be directly inserted into the convolutional neural network. This method can directly learn the segmentation task and automatically detect the ventricle contours from the original MRI images. Our method was evaluated using a quantitative comparison to existing state-of-the-art methods on the MICCAI 2009 left ventricle and 2012 right ventricle segmentation datasets. Excellent segmentation performance was observed with respect to common metrics used on both datasets. The average execution time of our method for predicting a slice was in milliseconds in the test phase. Our experiments demonstrated the efficacy of the proposed method based on attention mechanism for accurate and fast ventricle segmentation, indicating its potential clinical application.