Comparative analysis of U-Net and TLMDB GAN for the cardiovascular segmentation of the ventricles in the heart

Abstract

ObjectiveMagnetic Resonance Image (MRI) is an important imaging modality for diagnosing heart disease and analyzing heart function. The size and shape of the ventricle are important parameters for judging whether the heart is normal, and the ventricles in the MRI image is effectively segmented It is the key to obtain the ventricle size, shape and other parameters. Accurate segmentation of the entricle is the fundamental guarantee for the evaluation of cardiac function. However, in the heart image, the contrast between the ventricle area and the background area is not obvious, the boundary is blurred, and there is noise in most of the images. The accurate segmentation of the ventricle becomes a challenging problem. MethodsWe performed scanning of short-axis cardiac MR image sequences based on 33 subjects. Each subject has 8 to 15 sequences, each pertaining to a 20-frame sequence. Based on the U-Net neural network structure, the high-resolution information directly transferred from the encoder to the same-height decoder through the connection operation can provide more refined features for segmentation, such as gradients. The MRI left ventricular image segmentation method based on transfer learning and multi-scale discriminant Generative Adversarial Network (TLMDB GAN) solves the problem of insufficient ventricular image data. ResultsAccording to the experimental results of TLMDB GAN and U-Net network on the data set, the Dice coefficients of TLMDB GAN segmentation of the inner cardiac wall and outer cardiac wall of the ventricle are 0.9399 and 0.9697, respectively, which are 0.01 higher than other methods. The Dice coefficients of U-Net segmentation of the inner cardiac wall and outer cardiac wall of the ventricle are 0.8829 and 0.9292, respectively; ConclusionThe experimental results show that the TLMDB GAN based on transfer learning and multi-scale discrimination significantly improves the segmentation accuracy when compared with the U-Net segmentation model.