The complexity of the liver region in abdominal images and the limitations of traditional segmentation methods in feature extraction make the field of liver segmentation still have many challenges. Aiming at the shortcomings of the existing segmentation networks in the global and local information processing of the liver region, this paper proposed a segmentation method of cyclic densely connected networks that integrated more local features. This method integrated the cyclic dense connection module and the local feature supplement module into the learning unit of the coding process, so that the coding unit integrated deep-level global information and local feature information on a larger scale. Finally, after the decoding process, this method used the softmax function to output the segmentation result. On LiTS data set, the method performed well in multiple evaluation indicators, with an accuracy of 95.1%. In addition, related experiments on Data_67 dataset also proves that the method had good generalization performance. Experiments show that dense connections and more local information can make the performance of the liver segmentation model better.

[2021-08-05] Printed Article /doc/2020.08.0391/v2