Automatic Liver Segmentation Method from CT Images Based on Improved 3D U-Net

• Main Authors: Zhou, Yingxin, Zong, Jiashuai
• Format: Conference Proceeding
• Language: English
• Subjects: attention mechanism; Computed tomography; Convolution; CT image; deep supervision; Fuses; Image segmentation; liver segmentation; Measurement; neural network; Three-dimensional displays; Training
• ISSN: 2693-2865
• DOI: 10.1109/ITAIC54216.2022.9836869

Surgery is the best way to treat liver cancer, automatic and precise segmentation of the liver from the patient's abdominal CT image plays an important role in guiding the surgical procedure. It is difficult to segment the liver, not only because of the similar grayscale values and the low contrast between the liver and surrounding organs, but also the quite different shape of the liver in CT images between different individuals. To address the above problems, an automatic liver segmentation network based on multi-scale feature fusion (MSFF-Net) was proposed. The model was based on the basic network structure of 3D U-Net. At first, the residual dilated convolution module was used to expand the receptive field and obtain multi-scale information. In addition, the residual attention module was used to fully fuse the multi-scale features in the upper and lower networks to improve the network's attention to the detailed information of the target region. Finally, a hybrid loss function was designed for the data label imbalance problem, which was combined with deep supervision to improve the segmentation effect. The ablation study was conducted on the public datasets LiTS 2017 and 3DIRCADb. The model has performed on the LiTS 2017 dataset, Dice per Case and Dice Global metrics are 0.963 and 0.964 with an increase of 3.4% and 3.1% higher compared with the baseline network respectively, and the remaining auxiliary evaluation metrics were also relatively high. The Dice per Case and Dice Global evaluation metrics on the 3DIRCADb dataset were 0.961 and 0.961 with an increase of 3.3% and 3.1% higher compared with the baseline network respectively, which further demonstrates the effectiveness and generalization ability of the proposed method.