Liver cancer ranks as the sixth most commonly diagnosed cancer and the third leading cause of cancer-related death worldwide, with hepatocellular carcinoma (HCC) accounting for roughly 90% of all liver cancer cases. According to the Global Burden of Disease Study 2021, there were over 529,000 new cases and 483,000 deaths attributed to liver cancer in 2021 alone. Over the past two decades, incidence has increased by 53.7% and mortality has risen by 48.0%, reflecting a rapidly growing global health challenge.
Early-stage HCC often remains asymptomatic, leading to late detection when disease has already progressed to an advanced stage, limiting therapeutic options and worsening prognosis. The molecular and clinical heterogeneity of liver cancer further complicates treatment decisions, which may include surgical resection, liver transplantation, locoregional therapies (TACE), and systemic treatments (sorafenib, immunotherapy). These complexities underscore the need for advanced AI-driven tools.
This review, published in the International Journal of General Medicine by Zhang et al. from Heilongjiang University of Chinese Medicine, presents a unified AI pipeline for HCC that spans the full clinical workflow: data ingestion and preprocessing, tumor detection and diagnosis, personalized treatment planning, drug discovery, patient management, and research. The pipeline integrates multimodal data including imaging, genomic profiles, and clinical records, applying machine learning (ML) and deep learning (DL) techniques at each stage.
The core AI technologies covered include convolutional neural networks (CNNs) for image analysis, computer vision for automated tumor detection and segmentation, and natural language processing (NLP) for extracting actionable insights from electronic health records (EHRs) and biomedical literature. The authors argue that together, these technologies have the potential to transform the landscape of HCC diagnosis, treatment, and research.
CNN-based tumor detection: Convolutional neural networks are the backbone of AI imaging analysis in HCC. Trained on large datasets of CT, MRI, and ultrasound images, CNNs learn to automatically extract features through layered convolutional operations. A study by Said et al. evaluated CNNs for semiautomated HCC segmentation on MRI in 292 patients, finding strong results in single-slice segmentation on diffusion-weighted imaging (DWI) and precontrast T1-weighted imaging (T1WI pre) sequences, with accuracy correlating to tumor size.
Specific CNN variant performance: The review presents a detailed comparison table of architectures. DCNN-US on ultrasound achieved accuracy 84.7%, sensitivity 86.5%, specificity 85.5%, AUC 0.924, offering a cost-effective alternative to contrast-enhanced CT. DCCA-MKL on contrast-enhanced ultrasound (CEUS) reached accuracy 90.4%, sensitivity 93.6%, specificity 86.9%, AUC 0.953 with superior multi-phase image analysis. A two-input CNN on CT with integrated tumor marker data achieved accuracy 61.0%, sensitivity 75.0%, specificity 88.0%, AUC 0.870. Extremely Randomized Trees on MRI differentiated five lesion types with accuracy 88%, sensitivity 75.0%, specificity 56.0%, AUC 0.910.
U-Net for segmentation: The U-Net architecture, designed specifically for medical image segmentation, excels at delineating the liver and tumors from CT and MRI. A U-Net variant with residual connections demonstrated segmentation accuracy ranging from 0.81 to 0.93 on annotated CT datasets. The Successive Encoder-Decoder (SED) framework, consisting of two encoder-decoder networks in series, achieved a Dice score of 0.92 for liver segmentation and 0.75 for tumor prediction, with the ability to reconstruct 3D images from individual CT scans.
DeepLab V3+ for prognostic assessment: In a study of 209 patients, the DeepLab V3+ model achieved high segmentation accuracy on MRI and, when integrated into a decision fusion model, improved microvascular invasion (MVI) prediction with AUC 0.968. Combined with a nomogram, it enhanced early recurrence prediction (AUC 0.69 on validation) after surgery. The review concludes that U-Net and SED are best for treatment planning requiring anatomical delineation, while DeepLab V3+ is ideal for outcome prediction tasks like MVI.
The screening gap: Traditional HCC screening relies on ultrasound, which has suboptimal sensitivity, and serum alpha-fetoprotein (AFP), which remains negative in nearly two-thirds of early-stage HCC patients. These limitations mean that many high-risk patients, particularly those with chronic hepatitis B, hepatitis C, or cirrhosis, are not diagnosed until the cancer has advanced. AI addresses this by integrating demographic information, medical history, and laboratory results to identify high-risk individuals more effectively.
Deep learning on B-mode ultrasound: A deep learning model called Xception, trained on B-mode ultrasound images, achieved an AUC of 93.68% in identifying AFP-negative HCC among hepatitis B virus-infected patients, outperforming other architectures including MobileNet and ResNet50. This model demonstrated sensitivity of 96.08% and specificity of 76.92%, making it particularly valuable for the large population of patients whose HCC would be missed by AFP testing alone.
Gradient-boosting for chronic hepatitis B: A gradient-boosting machine algorithm developed specifically for chronic hepatitis B patients demonstrated superior predictive performance with a c-index of 0.79, outperforming traditional risk scores such as PAGE-B and REACH-B. The model identified a minimal-risk group with less than 0.5% HCC risk over 8 years, suggesting these patients could potentially undergo less intensive surveillance, reducing healthcare burden without compromising safety.
LiverColor tool: A tool called LiverColor uses image analysis and machine learning to assess hepatic steatosis (fatty liver), achieving 85% accuracy and an ROC curve of 0.82, outperforming traditional assessment methods. These AI-driven screening approaches not only improve diagnostic accuracy but also enable personalized risk stratification, allowing clinicians to tailor surveillance intensity to individual patient risk profiles.
Genomic and proteomic biomarkers: AI algorithms, including machine learning models and neural networks, are being applied to high-dimensional gene expression and proteomic data to uncover patterns associated with HCC. Haoran et al. studied microvascular invasion (MVI) in HCC using multi-transcriptomics data and identified a malignant cell subtype linked to MVI that was enriched in the MYC pathway and interacted via the MIF signaling pathway. Their newly developed prognostic model, based on MVI-related genes, showed superior accuracy compared to existing models for HCC management.
Liquid biopsy with ctDNA: While tissue biopsy remains the gold standard for mutational profiling, AI-enabled liquid biopsy approaches using circulating tumor cells (CTCs) and circulating tumor DNA (ctDNA) are showing promise. In a cohort study of 121 advanced HCC patients, mutational analysis of ctDNA revealed alterations in key HCC-associated driver oncogenes and tumor suppressor genes, including TERT promoter, TP53, PTEN, ARID2, KMT2D, and TSC2. This enabled identification of predictive mutational signatures associated with response to systemic therapy using tyrosine kinase inhibitors.
Limitations and complementary role: Despite the value of ctDNA as a tumor biomarker, the authors note important limitations: insufficient early detection sensitivity, lack of standardized protocols, and inability to capture tumor spatial heterogeneity. AI-based ctDNA profiling currently serves as a complementary tool for unresectable cases rather than a biopsy replacement. Multiparametric approaches combining ctDNA with imaging and clinical features are needed, and large-scale prospective validation trials remain essential before clinical adoption.
Personalized treatment planning: Traditional HCC treatment strategies rely on general protocols based on BCLC staging and Child-Pugh scores, but these may not be optimal for every patient given the heterogeneous nature of liver cancer. AI addresses this by integrating genomic, imaging, and clinical data to predict individual treatment responses. Machine learning algorithms analyze genetic profiles from next-generation sequencing to identify mutations associated with sensitivity or resistance to targeted therapies (tyrosine kinase inhibitors) or immune checkpoint inhibitors (immunotherapy), enabling more precise therapy selection.
Imaging-genomic integration: Deep learning algorithms analyze pre-treatment imaging scans to assess tumor characteristics including size, shape, and texture. By combining these imaging insights with genomic data and clinical histories, AI models predict which treatment options are most effective for each individual patient. This integrated approach helps avoid less effective treatments and reduces the likelihood of adverse side effects, improving quality of life alongside clinical outcomes.
Radiation therapy optimization: Deep learning algorithms analyze 3D imaging data from CT or MRI to create detailed anatomical models, optimizing radiation dose distribution based on tumor size, shape, location, and surrounding healthy tissues. A recent advancement combined a Gaussian filter with the nnU-Net architecture for liver tumor segmentation. Utilizing 130 cases from the LiTS2017 dataset for training and validation, the model achieved average Dice similarity coefficients (DSC) of 0.86 and 0.82 for validation and public testing sets. In clinical testing on 25 cases, the DSC improved from 0.75 to 0.81 after applying the Gaussian filter, demonstrating enhanced accuracy in distinguishing tumors from cysts.
Adaptive treatment: Treatment plans can be adjusted dynamically as therapy progresses, which is particularly beneficial when tumors shrink or change shape during treatment. By continuously updating imaging data, the system ensures that radiation remains accurately targeted throughout, reducing side effects and improving overall treatment effectiveness.
Preoperative planning with 3D models: Precise surgical planning is essential for liver surgeries where the goal is to remove HCC tumors while preserving healthy tissue. AI utilizes 3D imaging data to create detailed virtual models of the liver and tumor, enhancing visualization of tumor location relative to critical structures like blood vessels and bile ducts. These models enable surgeons to plan resections more accurately before entering the operating room.
Automated LI-RADS classification: Khaled et al. conducted a feasibility study to automate the Liver Imaging Reporting and Data System (LI-RADS) for HCC diagnosis using deep learning. They trained a U-Net-based deep CNN on multiphasic MRI data from 174 patients to automatically segment the liver and detect HCC. The model achieved high accuracy, with mean Dice similarity coefficient (DSC) of 0.91 for liver segmentation and 0.68 for HCC lesion detection, and reduced false positives through postprocessing. This approach suggests a more efficient and clinically practical implementation of the LI-RADS classification system.
Robotic intraoperative guidance: During surgery, robotic systems integrate real-time imaging with preoperative models to guide surgical instruments with high accuracy. Real-time feedback and predictive analytics from intraoperative data help identify potential issues such as excessive bleeding or changes in tumor position. This facilitates precise incisions and complex maneuvers, reducing the risk of complications and supporting more effective surgical decision-making.
Virtual screening: AI-driven virtual screening uses machine learning algorithms to predict interactions between small molecules and target proteins associated with HCC. Deep learning models analyze large chemical compound databases to identify potential drug candidates that might effectively interact with specific cancer-related targets, such as those involved in tumor progression signaling pathways. This dramatically accelerates the initial phases of drug discovery compared to traditional laboratory-based screening.
Molecular modeling and docking: AI algorithms generate detailed 3D models of drug molecules and their target proteins, allowing researchers to visualize how compounds fit into active sites. AI-driven molecular docking simulations predict binding affinities and identify which molecular modifications might enhance efficacy or reduce side effects, helping prioritize the most promising candidates for further experimental development and testing.
MDeePred for drug repurposing: The machine learning tool MDeePred leverages datasets from the Open Targets Platform, UniProt, ChEMBL, and Expasy databases to predict drug-target interactions (DTIs). Through enrichment analyses, MDeePred identified 6 out of 380 DTIs as promising candidates for HCC treatment. These candidates exhibited favorable drug-like properties including physicochemical characteristics, lipophilicity, water solubility, and medicinal chemistry profiles comparable to approved HCC drugs such as lenvatinib, regorafenib, and sorafenib. Molecular docking studies confirmed the binding efficacy of these compounds to HCC-associated targets.
miRNA regulatory networks: Researchers constructed a miRNA-gene regulatory network and used a shortest path-based method to assess the impact of miRNAs on cell fate genes. Results from breast and liver cancer datasets confirmed that differentially expressed miRNAs significantly affected cell fate genes governing the balance between cell proliferation and apoptosis, offering insights for potential therapeutic applications and drug sensitivity prediction in HCC.
Data quality and standardization: AI models rely on diverse data sources including imaging, genomic, and clinical records, but these often suffer from heterogeneity in formats, protocols, and standards. Imaging data from different machines or institutions varies in resolution, contrast, and acquisition techniques, complicating unified dataset creation. Clinical data from EHRs may be inconsistent due to variations in data entry practices, incomplete records, and missing values. Advanced imputation techniques add complexity and may impact model performance.
Bias and health disparities: Bias often stems from imbalanced datasets predominantly from development regions, which underrepresent certain ethnic groups, socioeconomic classes, and geographical areas. Hepatitis B virus-related HCC is more prevalent in Asian populations, while hepatitis C virus and alcohol-related HCC dominate in Western countries, leading to disparities in diagnostic accuracy and treatment recommendations. Socioeconomic disparities in access to advanced imaging further exacerbate bias, as models trained on data from tertiary centers may underperform in resource-limited settings. Addressing these issues requires diverse datasets, fairness-aware machine learning, and rigorous external validation.
Distributed learning and explainability: To enhance model robustness, distributed learning techniques like cyclical weight transfer (CWT) and fairness-aware methods are critical. Federated learning addresses data privacy concerns by training models across institutions without sharing raw patient data. Explainable AI (XAI) ensures transparency so clinicians can understand and trust model recommendations. Regulatory frameworks by agencies like the FDA and European Commission are evolving but remain complex, particularly for adaptive AI models that require ongoing oversight.
Future directions: The most promising paths forward include multimodal data integration combining imaging, genomic, and clinical records for improved accuracy; edge computing for real-time analysis on wearable devices to detect early disease progression; and advanced algorithms including deep learning, reinforcement learning, and transfer learning to improve model robustness. Collaborative AI systems integrating inputs from multiple institutions can enhance collective knowledge, while robust ethical and regulatory frameworks are essential to ensure transparency, accountability, and fairness in clinical deployment.