MShNet: Multi-scale feature combined with h-network for medical image segmentation

Objective: Medical image segmentation is the key foundation of medical image analysis. However, the uncertainty of the size, shape and location of the lesion greatly affect the segmentation accuracy. To solve the above problems, the multi-scale feature combined with the h-network (MShNet) is proposed in this paper. Methods: Firstly, a network framework, which is similar in shape to the letter "h" and consists of an encoder and two decoders is built to obtain stronger feature expression ability. The first decoder is responsible for obtaining the preliminary segmentation information of the image, and the second decoder enhances the feature expression of the nodule by fusing the information learned by the first decoder. Secondly, an enhanced down-sampling module is constructed in the encoder to reduce the information loss caused by down-sampling. In addition, to further reinforce the generalization ability of the model, the fusion convolutional pyramid pooling is designed to realize multi-scale feature fusion. Results: In the internal dataset of thyroid nodules, the DSC is 0.8721 and the HD is only 0.9356; DSC in the public dataset (DDTI,TN3K,ISIC and BUSI) also reached the optimal levels of 0.7580, 0.7815, 0.8853 and 0.7501 respectively and the HD for the last segmentation (Kvasir-SEG) is 16.5197. Conclusion: A large number of experimental results show that MShNet effectively improves the segmentation performance with less parameters, and achieves the most advanced performance in robustness and efficiency. Significance: The proposed algorithm provides a deep learning segmentation procedure that can segment thyroid nodule in ultrasound images effectively and efficiently.