Liver tumor segmentation using a new asymmetrical dilated convolutional semantic segmentation network in CT images

Abstract Liver cancer has become one of the commonly occurring cancers in men and women. The segmented computed tomography (CT) image of the tumor is the major source for a doctor to diagnose liver diseases. This article proposes a fully automatic system based on an asymmetric dilated convolutional encoder–decoder neural (ADCEDN) network for the problem of accurately extracting the liver and tumor from abdominal CT slices. Specifically, two ADCEDN networks are used in a cascaded form to perform both liver and liver abnormality segmentation. We enrich the encoder module by incorporating the multiscale contextual information through dilated convolutions that enhance the capacity and efficiency of the network. Hybrid dilated convolution across depths perform depth‐wise separable convolution by providing a less steep increase in dilation rate and thus denser sampling that can reduce the computation cost and several metrics while maintaining better performance. Also, the encoder–decoder network is made asymmetric, thereby making the network less complex and more generalizable. The suggested method produced an average dice similarity coefficient score of 96.54% and 76.16% for liver segmentation and liver abnormality segmentation on the 3D Image Reconstruction for Comparison of Algorithm Database data set and 96.42% and 75.7% on the Liver Tumor Segmentation data set. The test outcomes show that the proposed ADCEDN attains state‐of‐the‐art performance in both segmentation of liver and tumors.