Hepatocellular carcinoma (HCC) - the most common type of liver cancer - is the fifth most common cancer worldwide and the third-leading cause of cancer death. With approximately 500,000 new cases per year globally, it carries a grim five-year survival rate of just 15%. Most HCC tumors arise in livers already damaged by cirrhosis, caused by fatty liver disease, alcohol, hepatitis B, or hepatitis C.
Despite advances in treatment, including the drug combination Atezolizumab plus Bevacizumab for advanced HCC, outcomes remain poor largely because most cases are diagnosed at late stages. AI has recently emerged as a promising tool to improve all three phases of HCC management: predicting who will develop HCC before it appears, detecting it accurately when it does, and predicting outcomes and treatment response once diagnosed.
Several AI models have been trained on electronic health record data to predict which high-risk patients - those with cirrhosis or chronic viral hepatitis - will go on to develop HCC. These models analyze large volumes of longitudinal clinical data including lab values, demographics, and diagnoses over time. Results are promising: a recurrent neural network model in patients with chronic hepatitis C achieved an AUROC of 0.759 for predicting who would develop HCC, outperforming conventional statistical approaches.
One particularly useful potential application is personalizing surveillance strategies. Currently, all patients with cirrhosis undergo routine ultrasound scans every six months regardless of their individual risk level. AI-based risk scores could identify the specific subgroup of patients at highest short-term risk, allowing more intense monitoring to be focused where it matters most, while reducing burden on lower-risk individuals and on healthcare systems with limited capacity.
Standard surveillance for liver cancer uses B-mode ultrasound, but this test misses 37-54% of tumors. Deep learning models applied to ultrasound images can substantially improve accuracy. One large multicenter deep convolutional neural network achieved an AUROC of 0.92 for distinguishing benign from malignant liver lesions, matching the diagnostic accuracy of experienced radiologists (both 76.0%) and approaching that of CT scanning (84.7%).
For CT and MRI, AI has been used to classify indeterminate liver lesions - nodules that imaging cannot clearly define as cancer or non-cancer. CNN-based models have achieved diagnostic accuracies of 84-86% on these challenging cases. One three-phase CT model achieved a median AUROC of 0.92 for distinguishing HCC from other lesion types. Automated liver tumor segmentation has also advanced through deep learning, with the best models achieving Dice scores of 0.96 for liver segmentation, though tumor-specific segmentation remains harder.
Beyond imaging, deep learning is being applied to digitized liver biopsy slides (histopathology). Neural networks trained on H&E-stained slides can distinguish HCC from adjacent normal tissue with AUC above 0.90, and can even predict underlying genetic mutations - with AUCs from 0.71 to 0.89 for specific recurrent mutations. One key finding is that combining a pathologist with an AI model outperformed either alone, suggesting AI should augment rather than replace expert pathology review.
At the molecular level, multi-omics approaches combine RNA sequencing, DNA methylation, and protein data to identify biological subtypes of HCC. Deep learning-based analyses of RNA-seq data have identified two HCC subpopulations with significantly different survival rates, validated across five external cohorts (C-statistic 0.67-0.82). Single-cell RNA sequencing - which profiles thousands of individual cells simultaneously - is revealing the immune landscape of liver cancer in unprecedented detail, identifying exhausted T cells and tumor-associated macrophages that may be targets for immunotherapy.
For patients with established HCC, AI models are being used to predict survival and treatment response. Deep learning models applied to surgical pathology slides have predicted post-resection survival with higher accuracy than conventional scoring systems that use clinical and pathological features. One model using histology slides correctly identified that high-risk tumor tiles were enriched for features like a particular aggressive subtype and cellular abnormality - confirming the model was learning clinically meaningful patterns.
For patients receiving transarterial chemoembolization (TACE) - a procedure that delivers chemotherapy directly into liver tumor blood vessels - AI models using CT scan features have achieved AUCs of at least 0.94 for predicting whether patients will respond completely, partially, or not at all. Accurate prediction of who will benefit from TACE before the procedure could prevent patients who will not respond from undergoing an invasive treatment with side effects.
Despite the breadth of promising findings, clinical implementation of AI for HCC remains rare. Key barriers include the lack of standardized methods for AI data analysis, no universal approach to handling missing data in large datasets, and the fact that most studies are retrospective - looking back at historical data rather than testing AI in real-time clinical care. Performance of models trained retrospectively typically declines when tested prospectively on real-world data.
Data diversity is another critical concern. Most AI models for HCC have been trained on datasets that lack sufficient racial, ethnic, and socioeconomic diversity. Since AI model accuracy depends fundamentally on the characteristics of its training data, a model trained mainly on Asian or North American populations may perform less well in other regions. Data sharing between institutions and countries - including sharing of imaging data and molecular data from clinical trials - is identified as both an ethical and scientific imperative, though cultural, legal, and financial barriers remain.
The future of AI in HCC lies in prospective clinical trials where AI outputs are actively used to guide treatment decisions. Several such trials are already underway, including one evaluating an AI CT diagnosis algorithm against standard LI-RADS criteria and another building a risk stratification algorithm from ultrasound and clinical data. These trials will provide the rigorous evidence needed to determine whether AI tools truly improve patient outcomes.
Explainability - understanding why an AI model makes a particular prediction - is essential for clinical adoption. Doctors are more likely to trust and act on AI recommendations when they can see which imaging features or biological signals drove the prediction. Advances in 'explainable AI' that make neural network decisions transparent are a necessary complement to improving raw model accuracy. Together, standardization, data sharing, diversity, prospective validation, and explainability form the roadmap for AI to fulfill its potential in liver cancer care.