Improved SwinUNet with fusion transformer and large kernel convolutional attention for liver and tumor segmentation in CT images
Linfeng Jiang, Jiani Hu, Tongyuan Huang
Abstract
Segmentation of both liver and liver tumors is a critical step in radiation therapy of hepatocellular carcinoma. Despite numerous algorithms have been proposed for organ and tumor delineation, automatic segmentation of livers and liver tumors remains a challenge due to their blurred boundaries and low tissue contrast compared to surrounding organs within CT images. The U-Net-based methods have achieved significant success in this task. However, they often suffer from the limitation that feature extraction lacks relationships, i.e., context, among adjacent areas, thereby leading to uncertainty in segmentation results. To address with this challenge, we incorporate both global-local context and attention into the Swin-UNet. Firstly, we introduce a Swin-neighborhood Fusion Transformer Block (SFTB) to capture both global and local context in an image, enabling us to distinguish instances and their boundaries effectively. Secondly, we design a Large-kernel Convolutional Attention Block (LCAB) with two types of attention to highlight crucial features. Experiments on the LiTS and 3D-IRCADb datasets demonstrate the effectiveness of the proposed method, with dice scores of 0.9559 and 0.9610 for liver segmentation, and 0.7614 and 0.7138 for liver tumor segmentation. The code is available at https://github.com/JennieHJN/image-segmentation/tree/master .