Radiomics-guided GAN for Segmentation of Liver Tumor Without Contrast Agents

Segmentation of the liver tumor is critical for preoperative planning, surgical protocol guidance, and post-operative treatment. Because of using contrast agents (CA), current liver tumor imaging still suffers from high-risk, time-consumption and expensive issues. In this study, a new Radiomics-guided generative adversarial network (Radiomics-guided GAN) is proposed as a safe, short time-consumption and inexpensive clinical tool to segment liver tumor without CA. The innovative Radiomics-guided adversarial mechanism learns the mapping relationship between the contrast images and the non-contrast images, which leads to completing the segmentation. Radiomics-guided GAN contains a segmentor and a discriminator module: the discriminator innovatively uses the Radiomics-feature from the contrast images as prior knowledge to guide the segmentor's adversarial learning; the segmentor innovatively uses dense connection and skip connection to receive and share the guidance information, extracting the representing feature - Implicit Contract Radiomics (ICR) feature - in the non-contrast images. Our method yielded a pixel segmentation accuracy of 95.85%, and a Dice coefficient of 92.17 +/- 0.79%, from 200 clinical subjects. The results illustrate that our method achieves the segmentation of liver tumor without CA and become the most potential useful tool for clinicians.