Hybrid attention mechanism for liver tumor segmentation in CT images

Abstract

A new method for automatic liver tumor segmentation from computed tomography (CT) scans based on deep neural network is presented. A modified cascaded U-Net with hybrid attention mechanism is designed to segment the liver and liver tumors respectively. The embedded hard attention mechanism in the deep neural network allows the network to automatically interpret the input image to obtain more effective information, and it is end-to-end trainable along with the cascaded U-Net. At the same time, the joint channel attention and spatial attention mechanisms enhance extraction of effective features. The Liver Tumour Segmentation (LiTS) dataset is used to evaluate the relative segmentation performance obtaining an average dice score of 0.762 using the new method.