Improved 3D fully convolutional network based on squeeze-excitation method for liver segmentation

Due to the variable shape and size of the liver in abdominal CT and the complex surrounding tissues, accurate liver segmentation in CT remains a challenge. In the field of two-dimensional image analysis, the squeeze-excitation (SE) method effectively adjusts the input feature information. Our work is based on SE method to obtain feature information that is more relevant to the liver region through improved feature recalibration methods. This work adds the MSCR block, which can do space and channel feature recalibration, to 3D full convolution network. The 3D space part of the MSCR block re-calibrates the feature of liver voxel space. And in the channel recalibration part, it changes the original spatial compression operation according to the none-local method, so the spatial sampling results contain more global context information. Multiple metrics are used to evaluate the proposed model on the LiTS dataset, and it achieves better segmentation performance than other comparison models such as Attention U-Net. The method is further tested on the 3DIRCADb dataset, which proves its effectiveness and stability. Thus, the model proposed in the paper is effective for improving the performance of liver segmentation.