Deep learning for automatic segmentation of hepatocellular carcinoma in contrast enhanced CT scans

Liver cancer represents a significant cause of cancer-related mortality, with hepatocellular carcinoma (HCC) being the most prevalent forms. Computed tomography (CT) serves as the principal imaging modality for the diagnosis of liver tumors, particularly HCC. The precise identification of tumor presence and location necessitates highly skilled radiologists. Consequently, automated liver tumor segmentation from CT images offers a valuable tool to support cancer diagnosis and treatment planning. Nevertheless, this task presents considerable challenges due to the inherent variability in tumor shape, dimension, and imaging acquisition techniques. In this paper, we evaluate state-of-the-art segmentation architectures across a range of datasets. These include the publicly accessible Liver Tumor Segmentation (LiTS) dataset, which covers a spectrum of liver lesions, as well as the HCC-TACE-Seg and WAW-TACE datasets, comprising CT scans of HCC patients prior to treatment. In addition, we introduce a novel dataset of contrast-enhanced CT (CECT) scans that are routinely used for HCC diagnosis. The focus of this manuscript is on the segmentation of both the liver and tumors, with specific attention directed toward HCC. This study provides a comparative analysis of segmentation models applied to diverse datasets, encompassing both public and proprietary datasets.