Automatic Liver Tumor Segmentation from CT Images with Dual Attention Mechanism

Liver cancer is one of the most common malignant tumors, and it is a serious threat to human life and health. Accurate segmentation of liver tumors from computed tomography (CT) images is crucial in the diagnosis and treatment of liver tumors. Due to the complexities of liver tumors such as their varied morphology, sizes, significant density variations, and blurred boundaries, accurate segmentation of liver tumors remains an extremely challenging task. This paper utilizes the U-Net as the backbone network and constructs a 2.5D convolutional neural network to enhance the efficiency of spatial information utilization. By introducing dual attention (DA) modules composed of the coordinate attention and the spatial attention, the model can more effectively focus on target regions in the image. Moreover, through the Ladder Atrous Spatial Pyramid Pooling (Ladder-ASPP) module, the model merges high-level multi-scale features to obtain and utilize multi-scale spatial context information. Additionally, a hybrid loss function combining binary cross entropy (BCE) and Dice is employed to speed up the network's convergence. The proposed 2.5D network is trained and tested using the LiTS2017 dataset. Experimental results demonstrate that the proposed 2.5D dual attention network includes Ladder-ASPP module (DAL-UNet) achieves a Dice coefficient of 73.63%, outperforming the current methods.