Automatic Cardiac MRI Segmentation Using Deep Learning

In the biomedical realm, automatic medical image segmentation is regarded as a challenging research topic. Among the different automatic methods, U-shaped models have significantly advanced a wide range of medical image segmentation. When training deep neural networks like U-net, gradient degradation is one of the problems we need to prevent from occurring. In addition, giving more focus to the important objects in the image while disregarding unneeded areas is one of the desired properties in medical image segmentation. In order to prevent gradient degradation and focus on the important objects in the image, a novel U-net-based architecture is proposed for cardiac MRI segmentation. The model uses the merit of U-Net, residual U-Net (ResU-Net), and Attention U-Net (AttU-Net) for better performance and prediction results. U-Net, ResU-Net, and AttU-Net are used as benchmark models. For the model to efficiently train the data and to improve its accuracy, a data preprocessing mechanism is applied. The data preprocessing step includes contrast enhancement, data augmentation, and data normalization. For the contrast enhancement method, we tested three different contrast enhancement methods. These methods are non-linear contrast enhancement (gamma correction), multi-scale Retinex, and adaptive histogram equalization. After testing the three contrast enhancement methods on the data, the best one was selected for use in the deep learning algorithms. The proposed and benchmark models are tested on data from Anzhen Hospital and ACDC public data set. The result shows that the proposed model achieved a better prediction result on both data sets than the benchmarks. This is due to the benefits of using ResU-Net and AttU - Net, which enable it to concentrate on the small class, reintroduce features, and prevent gradient degradation, resulting in its advantage.