Compressed Channel Attention Mechanism for 3D Medical Image Segmentation of Liver

Abstract

AbstractMore and more Convolutional Neural Networks (CNNs) are used in computer vision, especially in the field of medical images. Since most of the medical data in clinical practice are three-dimensional, which is mainly obtained from imaging techniques such as MRI and CT, the use of the previous two-dimensional neural network becomes untimely. In this work, an end-to-end trained 3D image segmentation network combined with the Compressed Channel Attention Module (CCAM) is proposed to learn to predict the segmentation of the entire liver at one time. We embed the CCAM in the network structure of up-sampling and down-sampling. First, we divide the feature map obtained by down-sampling into two parts. The first part performs global average pooling and maximum pooling and then splicing, and the other part performs 1 × 1 convolution on the feature map. Then the two parts are merged, and finally spliced with the up-sampled feature map to realize the use of down-sampling features to monitor the up-sampled features, focusing on specific livers, and suppressing irrelevant areas in the input image. Experiments have been carried on the LITS dataset. Compared with the existing segmentation methods, the proposed method has better segmentation performance in both subjective and objective evaluation.Keywords3D image segmentationAttention mechanismDeep learning