AD-UNet: A Multi-Attention Mechanism Densely Connected U-Net for Liver Tumor Segmentation

As an important metabolic organ in human body, liver is also one of the most common sites for tumors. Since liver tumors are mostly malignant tumors, early screening of tumors is extremely crucial. To this end, liver tumor segmentation has contributed to improve the efficiency of liver cancer treatment and lowered the risk of human death. Traditionally, cancer diagnosis relies on manual judgment by physicians according to patients’ Computed Tomography (CT) images, which is inefficient and labor-intensive. On the other hand, since liver tumors are complex in distribution and varying in shape and size, neither traditional morphological image processing algorithms nor existing deep learning methods which suffer from inadequate feature extraction can extract liver tumor in a precise way. In this paper, we propose a multi-attention mechanism densely connected U-Net, AD-UNet, which combines dense connectivity and attention mechanism to strengthen the mapping of features and enhance the effective features. According to the experimental results, AD-UNet achieved very competitive result compared to other method. Efficacy of AD-UNet was demonstrated using the public dataset of Liver Tumor Segmentation (LiTS) Challenge 2017 and the 3D-IRCADb dataset. For liver tumor segmentation, AD-UNet achieved Dice of 85.4%, VOE of 22.9%, RVD of 16.9%, ASD of 0.927mm and MSD of 3.546mm with LiTS. And in the 3D-IRCADb dataset we obtained Dice of 67.27%, VOE of 37.63%, RVD of -0.82%, ASD of 1.427mm and MSD of 7.316mm.