Liver Cancer Diagnosis in CT Employing U-Net

Liver cancer is one of the most common types of cancer worldwide. Diagnoses and treatments in early stages of liver cancer can help to increase the survival rate. However, early treatments are based on early diagnoses, which require CT scanners as well as trained doctors to spot the tumor from CT images. The development of deep learning provides a possible way to fulfill such requirements. In this paper, a U-net based segmentation algorithm is proposed for liver tumor segmentation from CT images. Various numbers of filters were set to check their improvements of the model prediction accuracies along with three different types of loss functions. The results showed that mean squared error behaved better than the other two types of losses, and the number of filters makes no obvious contribution to the prediction accuracy. When the prediction accuracy rise, the actual behavior may decrease in CT scan with small tumor. A dice score of 0.91 was obtained with a model that can afford decent diagnosis results for both big and small tumor, which shows the potential of this method as a computer-aided diagnosis of liver tumor from CT images.