Background
In recent years, convolutional neural networks have been prominent in medicine image processing, but single convolution and frequent pooling operations very easily produce redundant information or miss key information. This paper designs a residual squeeze and excitation attention network (RSEA-Net) to solve the above problems.
Methods
The network has the following main advantages: (1) It can capture richer and more detailed features at different scales in our proposed step convolution module (SCM), which adopts a parallel structure design and has receptive fields of different sizes. (2) This paper also designs a residual squeeze and excitation attention module (RSEAM), which can improve the useful feature gain through space and channel. It can not only eliminate some redundant information but also improve the overall robustness of the model.
Results
This paper verifies the performance of the RSEA-Net in 2D lung CT and tongue image databases. This paper selects seven models for comparison. The experimental results showed that: The RSEA-Net has multi-size respective fields and can eliminate redundant information. In the 2D lung CT database, the Accuracy, Jaccard coefficient, and Dice coefficient reach 0.9939, 0.9705, and 0.9850, respectively. In the tongue image database, the Accuracy, Jaccard coefficient, and the Dice coefficient of the RSEA-Net reach 0.9954, 0.9794, and 0.9895, respectively, which are better than those of the other seven models.
Conclusion
We propose a new deep network model (RSEA-Net). The structure is composed of two U-shaped networks with left and right layers. These are precoding networks and precision segmentation networks. To avoid losing information or producing invalid information caused by frequent convolution and pooling operations, this paper designs an SCM to obtain more multi-scale features. This paper also designed an RSEAM, which can improve the useful feature gain through space and channel, remove some redundant information, and improve the network’s overall robustness. Finally, we also reduced the number of down-sampling operations and simplified the longitudinal complexity. Experimental results show that our method is superior to the original U-Net and other state-of-the-art methods in two different datasets.