Advancing Diagnostic Accuracy in Liver Cancer: A Systematic Review of Artificial Intelligence Applications in Hepatocellular Carcinoma and Cholangiocarcinoma Detection Using Abdominal CT Imaging.

OBJECTIVE

This study aimed to systematically evaluate the diagnostic performance of artificial intelligence (AI) in differentiating hepatocellular carcinoma (HCC) from cholangiocarcinoma (CCA) using abdominal CT and MRI, with an emphasis on its clinical implications for liver cancer management.

METHODS

Following the PRISMA guidelines, we conducted a comprehensive literature search across five major databases (PubMed, Web of Science, ScienceDirect, Scopus, and Google Scholar) from 2000 to May 6, 2025. Eligible studies included original research that applied AI for the diagnosis of HCC or CCA. Data were extracted on study design, population characteristics, imaging modality, AI methodology, diagnostic performance (sensitivity, specificity, accuracy, AUC), validation strategies, and risk of bias, which was assessed using QUADAS-2.

RESULTS

A total of 44 studies met the inclusion criteria. Most were retrospective, while only a few prospective designs provided real-time validation. CT and MRI were the dominant imaging modalities, with MRI showing superior sensitivity for small lesions, while CT was more effective for large tumors and vascular involvement. Convolutional neural networks (CNNs) were the most frequently used model architectures, although more advanced deep learning and hybrid radiomic-clinical models were also reported. Diagnostic performance was consistently strong: sensitivity and specificity ranged from 75% to 100%, overall accuracy from 73% to 96%, and AUC values from 0.74 to 0.99. Studies incorporating multi-modal imaging (CT+MRI) or radiomic-genomic features achieved the highest diagnostic performance, with accuracy and specificity exceeding 90-95%. Subgroup analyses revealed that tumor size, location, microvascular invasion, and patient demographics influenced AI model performance. Risk of bias was generally low-to-moderate, with limitations related to retrospective data and limited external validation.

CONCLUSION

AI models, particularly CNN- and radiomics-based, show accuracy comparable to radiologists in distinguishing HCC from CCA. Multi-modal integration and feature fusion hold the greatest promise for improving workflows. Large-scale, multi-center validation is needed to confirm their utility and enable adoption in liver cancer care.