Automated deep system for joint liver and tumor segmentation using majority voting

Abstract

In this paper, a system based on deep learning and majority voting is proposed for joint segmentation of the liver and hepatic tumors. The proposed system is composed of three steps. First, deep learning is utilized to extract deep features that describe the Computed Tomography (CT) images as well as cancerous nodules, using three different Convolutional Neural Networks (CNNs), i.e., VGG16-Segnet, Encoder-Decoder (ED)-Alexnet, and Resnet18. Second, a classification step using the extracted deep learning features is performed for each investigated network. To produce the final liver and hepatic tumor segmentation, the last step applies a majority voting technique to fuse the three utilized CNN outputs. To test the performance of the proposed system, the MICCAI LITS challenge database is used, composed of 130 CT volumes with a total of 16,917 cross-section images. The proposed system achieves Dice Similarity Coefficients (DSCs) of 94% and 76% for liver and lesion segmentations, respectively. Comparison with the related methods confirms the promise of the proposed system for joint liver and tumor segmentations