Abstract
BACKGROUND: Automatic segmentation of liver regions as well as liver lesions such as hepatocellular carcinoma (HCC) from computed tomography (CT) images is critical for accurate diagnosis and therapy planning. With the advent of deep learning techniques such as transformers, computer-aided diagnostic tools (CADs) have the potential to increase the accuracy of liver tumor diagnosis, progression, and treatment planning. However, two major challenges remain: 1) existing models struggle to extract robust spatial features for accurate liver and liver lesion segmentation, and 2) publicly available liver datasets with HCC annotations are limited. METHODS: We first present a new liver dataset acquired from Chongqing University Cancer Hospital (CCH-LHCC-CT) with HCC annotations. Second, we developed a novel deep learning architecture (ICT-Net), which is constructed based on a pretrained transformer encoder in conjunction with an advanced feature upscaling and enhanced convolution-transformer decoder formation. RESULTS: We performed liver and liver tumor segmentation on the CCH-LHCC-CT and three public CT liver datasets. The proposed ICT-Net architecture achieves superior accuracy (higher ACC/DSC/IoU, lower HD95) across all datasets. CONCLUSIONS: We construct a novel deep-learning architecture that produces robust information for liver and liver tumor segmentation. The statistical and visual results demonstrate that the proposed ICT-Net outperforms other existing approaches investigated in this study in terms of ACC, DSC, and IoU. Clinical Translation Statement: ICT-Net enhances surgical planning accuracy through precise tumor margin delineation and improves therapy response assessment reliability, which holds meaningful promise to support more precise and effective clinical therapeutic strategies for patients with HCC.