MAPFUNet: Multi-attention Perception-Fusion U-Net for Liver Tumor Segmentation

Abstract

The second-leading cause of cancer-related deaths globally is liver cancer. The treatment of liver cancers depends heavily on the accurate segmentation of liver tumors from CT scans. The improved method based on U-Net has achieved good performance for liver tumor segmentation, but these methods can still be improved. To deal with the problems of poor performance from the original U-Net framework in the segmentation of small-sized liver tumors and the position information of tumors that is seriously lost in the down-sampling process, we propose the Multi-attention Perception-fusion U-Net (MAPFUNet). We propose the Position ResBlock (PResBlock) in the encoder stage to promote the feature extraction capability of MAPFUNet while retaining the position information regarding liver tumors. A Dual-branch Attention Module (DWAM) is proposed in the skip connections, which narrows the semantic gap between the encoder's and decoder's features and enables the network to utilize the encoder's multi-stage and multi-scale features. We propose the Channel-wise ASPP with Attention (CAA) module at the bottleneck, which can be combined with multi-scale features and contributes to the recovery of micro-tumor feature information. Finally, we evaluated MAPFUNet on the LITS2017 dataset and the 3DIRCADB-01 dataset, with Dice values of 85.81 and 83.84% for liver tumor segmentation, which were 2.89 and 7.89% higher than the baseline model, respectively. The experiment results show that MAPFUNet is superior to other networks with better tumor feature representation and higher accuracy of liver tumor segmentation. We also extended MAPFUNet to brain tumor segmentation on the BraTS2019 dataset. The results indicate that MAPFUNet performs well on the brain tumor segmentation task, and its Dice values on the three tumor regions are 83.27% (WT), 84.77% (TC), and 76.98% (ET), respectively.