Automatic liver segmentation using U-net in the assistance of CNN
Yang Liu, Wei Yang · 2020
Liver segmentation is widely used in medicine for volumetric or morphological analysis. This paper presents a system to automatically segment liver from the CT scans for real clinical application. The proposed system uses one of the most suitable networks in biomedical segmentation U-net in the assistance of Convolutional Neural Network (CNN). The preprocessed CT scans with and without liver are divided into five position groups and labelled with group number. In the next step, we train CNN for automatic localization of the CT scans. Afterwards, three networks based on U-net are fed with images in the same group having liver and then trained for liver segmentation separately. The proposed system is evaluated on the scans from LiTS2017 and 3Dircadb database. The average IOU score improves to 92.8% compared with state-of-the-art. Meanwhile, the novel system demonstrates its flexibility in the choice of training data and its universality in testing data, which is suitable for both scans with liver and scans without liver. The experimental results indicate that the proposed system facilitates the procedure of liver segmentation and has its potential for accurate liver segmentation for clinical application.