MUNet: A Multi-scale U-Net Framework for Medical Image Segmentation

Artificial intelligence has once again become the focus of attention in all fields, of which deep learning has brought a series of changes in the field of computer vision. In this paper, we propose the MUNet model, which is a more general convolutional neural network framework for medical image segmentation. The framework proposed in this paper is essentially a fully convolutional encoder-decoder network based on feature pyramids, in which the encoder and decoder are connected by skip connections. Not only is it suitable for image segmentation, but it can also identify categories of regions of interest. The full convolutional neural network architecture can implement multi-scale image input and prediction. We compared the MUNet and UNet models in cervical lymph node localization and benign and malignant diagnosis in ultrasound images. Experiments show that on our dataset, MUNet with multi-scale segmentation has achieved a Dice score improvement of 4.1% and an AUC score improvement of 8.1% compared to U-Net.