DALU-Net: Automated Liver Segmentation and Volumetry for Liver Transplantation in Abdominal Computed Tomography Volumes
Abstract Liver transplantation is performed in patients with liver disease, using the liver of a braindead or living donor. In living-donor liver transplantation, the safety of the donor is critical. In addition, the amount that can be resected from the living donor is limited. It is important that accurately measuring the liver volume to avoid graft size mismatch. In this paper, we designed a deep attention convolutional long short-term memory (CLSTM) network architecture for liver segmentation that combines an attention mechanism, deep supervision, and CLSTM. The proposed model can focus on the liver in abdominal CT volume data and can learn inter-slice using CLSTM. Our framework was trained using 133 training cases, 29 validation cases, and 29 test cases of liver donors. We compared livers and volumes manually labeled by a liver transplant surgeon and those obtained by automatic segmentation of livers and volumes. We further evaluated the segmentation and volumetry of the left lobe, right lobe, and caudate lobe, according to the anatomical structure of the liver. Our approach significantly outperformed the 3D U-Net in terms of accuracy. Our approach can be used as an aid in estimating liver volume from CT volume data for living-donor liver transplantation.