Deep Joint Segmentation of Liver and Cancerous Nodules From Ct Images

A framework is proposed for joint liver and cancerous nodule segmentation from abdomen computed tomography (CT) images. The proposed framework consists of three main units. First, a preprocessing unit is used to enhance the image contrast. Second, two different deep convolutional-deconvolutional neural networks (CDNN), namely, Alexnet and Resnet18 models, are investigated to extract the features of liver images. Finally, a pixel wise classification unit is performed to provide the final segmentation maps of the liver and tumors. Results on the challenging MICCAI’2017 liver tumor segmentation (LITS) database, using Alexnet model and 4-fold cross-validation, achieve a Dice similarity coefficient of 90.4% for liver segmentation and of 62.4% for lesion segmentation. Comparative results with related techniques for joint liver and tumor segmentations show the effectiveness of the proposed framework.