MMNet: A multi-scale deep learning network for the left ventricular segmentation of cardiac MRI images

With the development of deep learning network models, the automatic segmentation of medical images is becoming increasingly popular. Left ventricular cavity segmentation is an important step in the diagnosis of cardiac disease, but post-processing segmentation is a time-consuming and challenging task. That is why a fully automated segmentation method can assist specialists in increasing their efficiency. Inspired by the power of deep neural networks, a multi-scale multi-skip connection network (MMNet) model is proposed to fully automate the left ventricular segmentation of cardiac magnetic resonance imaging (MRI) images; this model is simple and efficient and has high segmentation accuracy without pre-detecting left ventricular localization. MMNet redesigns the classic encoder and decoder to take advantage of multi-scale feature information, effectively solving the problem of difficult segmentation due to blurred left ventricular edge information and the low accuracy of end-systolic segmentation of the cardiac area. In the model encoding stage, a multi-scale feature fusion module applying dilated convolution is proposed to obtain richer semantic information from different perceptual fields. The decoding stage reconstructs the full-size skip connection structure to make full use of the feature information obtained from different layers for contextual semantic information fusion. At the same time, a pre-activation module is used before each weighting layer to prevent overfitting phenomena from arising. The experimental results demonstrate that the proposed model has better segmentation performance than advanced benchmark models. Ablation experiments show that the proposed modules are effective at improving segmentation results. Therefore, MMNet is a promising approach for the left ventricular fully automated segmentation.