Left ventricular full segmentation from cardiac Magnetic Resonance Imaging via multi-task learning

Heart disease has become the “number one killer” of human health, directly or indirectly endangering human health in the form of complications. Segmentation of the left ventricle can provide diagnostic value for evaluating cardiac health and identifying certain pathology. In this paper, we propose method called multi-constrained segmentation (MC-Seg) automatically calculates and measures the full segmentation results of left ventricle. We use U-Net convolution neural network architecture first to segment left ventricle as a benchmark. Taking into account the correlation between tasks, we add multitask learning to the segmentation results at the same time to improve the performance and improve the generalization of the network. Our method was validated and evaluated on the MICCAI 2018 ventricular dataset. After 50 rounds of training, by segmenting the cardiac MRI images from 145 patients, we reduced the training loss to 0.002 and the average dice overlap coefficients of the test set were 0.886. The results show that our method is innovative and can effectively improve the performance of left ventricular segmentation.