Robust Segmentation of the Left Ventricle from Cardiac MRI via Capsule Neural Network
Jiaxiu Dong, Chang Liu, Cong Yang, Nan Lin, Yangjie Cao · 2018
Segmentation of the left ventricle from cardiac magnetic resonance images provides important supplementary information for the diagnosis and treatment follow-up of cardiovascular diseases, the main cause of deaths worldwide. In this paper, we propose a novel capsule neural network to robustly and accurately extract myocardial borders of the left ventricle. The whole network consists of convolutional layers, primary capsule layer, digital capsule layer, fully connected layers, and de-convolutional layers. The output of digital capsule layer is generated by dynamic routing, enabling our method to cope with noise very well. The proposed method is trained and validated on the data from MICCAI 2013 left ventricle segmentation challenge. We also validate our method on images with different levels of noise. The proposed method exhibits desirable Dice coefficient on the origin data is 0.9417, and shows very little decrease even the images are noised heavily.