Artificial intelligence and radiomics on computed tomography for differentiating hepatocellular carcinoma and intrahepatic cholangiocarcinoma: a multimodal integration approach.
Authors
Affiliations (3)
Affiliations (3)
- Department of Hepatobiliary and Vascular Surgery, The First Affiliated Hospital of Chengdu Medical College, Chengdu Medical College, Baoguang Avenue 278, Xindu District, Chengdu, Sichuan Province, 610500, China.
- Department of Hepatobiliary and Pancreatic Surgery, Chengdu 363 Hospital Affiliated to Southwest Medical University, Chengdu, Sichuan Province, 610000, China.
- Department of Hepatobiliary and Vascular Surgery, The First Affiliated Hospital of Chengdu Medical College, Chengdu Medical College, Baoguang Avenue 278, Xindu District, Chengdu, Sichuan Province, 610500, China. [email protected].
Abstract
Differentiating between hepatocellular carcinomas (HCC) and intrahepatic cholangiocarcinomas (ICC) non-invasively is clinically critical but challenging, impacting patient management. This study aims to employ contrast-enhanced computed tomography (CECT) radiomics and artificial intelligence to distinguish HCC from ICC. This retrospective study (October 2014–July 2024) included 186 patients with pathologically confirmed HCC (<i>n</i> = 111) or ICC (<i>n</i> = 75) from two centers. A total of 134 patients from Center A were randomly divided into a training cohort (<i>n</i> = 93) and an internal validation cohort (<i>n</i> = 41), while 52 patients from Center B were assigned to an external validation cohort. For each patient, 4,520 radiomic features were extracted from the volume of interest. After feature reproducibility assessment, stable features were selected to construct the radiomics-based model. Logistic regression analysis was used to identify independent clinical predictors. A dataset comprising region of interest images per patient was analyzed using multiple neural network architectures. Outputs from the radiomics, clinical, and deep learning branches were fused via the fusion neural network to create the final model. Logistic regression identified two independent clinical predictors: cirrhosis status and elevated carbohydrate antigen 19–9. In radiomics-based module, the support vector machine model based on features from four CECT phases achieved an area under the curve (AUC) of 0.96 in a single internal validation assessment. In the neural network module, the self-designed hybrid neural network reached an AUC of 0.92 in the internal validation set. The fusion neural network achieved an AUC of 0.99 in the internal validation set and an AUC of 0.95 in the external validation set, outperforming the single-modality model. The fusion neural network shows potential in distinguishing HCC from ICC and may support early individualized treatment planning. Further validation with automated segmentation and explainable tools is warranted to confirm broader generalizability and facilitate clinical adoption. The online version contains supplementary material available at 10.1186/s12880-026-02305-3.