Introduction
Diagnostic imaging is an essential tool for the non-invasive detection and characterization of diseases, particularly liver malignancies. The accurate differentiation of liver cancers, such as hepatocellular carcinoma (HCC) and cholangiocarcinoma (CCA), remains a significant challenge, often requiring the expertise of highly trained radiologists. However, interpretation variability exists, especially among less experienced or newly certified radiologists, which can lead to diagnostic delays and inconsistencies [1].
To address these challenges, artificial intelligence (AI) has emerged as a transformative force in medical imaging, offering the potential to enhance diagnostic accuracy and efficiency. By mimicking human cognitive functions, such as reasoning and decision-making, AI enables machines to autonomously process complex data with minimal human oversight [2]. The adoption of AI in healthcare is rapidly accelerating, with advances in computing power and large-scale data analytics driving improvements in diagnostic workflows and treatment planning [3].
AI applications are now being extended beyond hospitals to support community-level health services, including public health data collection and disease surveillance.
Globally, liver cancer is the sixth most commonly diagnosed cancer and the third leading cause of cancer-related mortality, accounting for approximately 905,000 new cases and 830,000 deaths annually [4]. HCC constitutes about 75–85% of primary liver cancers, while CCA represents 10–15% [5]. The burden is particularly pronounced in Southeast Asia, including Thailand, Laos, and Cambodia, where CCA prevalence is among the highest worldwide, largely driven by liver fluke infection and chronic biliary inflammation [6, 7].
These epidemiological disparities underscore the pressing need for precise, accessible diagnostic tools tailored to both global and regional contexts.
Liver cancer diagnosis, which requires specialized radiologic expertise, stands to greatly benefit from AI advancements. Deep learning models, in particular, have demonstrated superior performance in the classification, detection, and characterization of liver lesions in imaging studies [8]. Several reports have indicated that AI systems can achieve diagnostic accuracies comparable to, or even exceeding, those of experienced radiologists, especially in identifying HCC and CCA using computed tomography (CT) and magnetic resonance imaging (MRI) [9, 10].
AI has shown remarkable capabilities in analyzing complex imaging data, such as multiphasic CT [11], and has outperformed conventional imaging techniques in certain contexts, such as mass spectrometry-based automated liver cancer diagnosis [12]. Furthermore, AI models are increasingly being used in radiomic analysis to predict the aggressiveness and treatment response of liver tumors [13].
Nevertheless, important challenges remain. Data heterogeneity, algorithm generalizability, and integration into real-world workflows limit translation into daily practice. Moreover, previous reviews have often been descriptive, with limited pooled statistical analyses and without structured risk-of-bias assessments such as QUADAS-2.
Few studies have addressed practical considerations such as interpretability, cost-effectiveness, and workflow implementation. These gaps highlight the need for more comprehensive evidence synthesis.
In light of these advancements and unmet needs, our research team undertook a systematic review to comprehensively evaluate the role of AI in the imaging-based diagnosis of hepatocellular carcinoma and cholangiocarcinoma. By incorporating pooled sensitivity, specificity, and AUC analyses, assessing study quality using QUADAS-2, and presenting visual comparative charts, this study aims to provide a robust overview of AI’s diagnostic potential. Additionally, we examine the predictive factors incorporated in AI models for liver lesion classification and discuss their practical implications for clinical adoption.
Ultimately, the goal of this review is to inform future AI development and enhance its clinical utility in diagnosing liver cancers, thereby contributing to improved disease prevention and control strategies [14-16].
Materials and Methods
This systematic review adhered to the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) guidelines [10]. A comprehensive literature search was conducted on May 6, 2025, across five major electronic databases PubMed, Web of Science, ScienceDirect, Scopus, and Google Scholar chosen for their broad coverage of biomedical and technical research. The search period was set from 2000 to 2025, reflecting the rise of AI applications in medical imaging during this timeframe.
Search Strategy
The search employed a combination of predefined keywords relevant to AI applications in liver cancer diagnosis, using Boolean operators to maximize comprehensiveness. The final search query included:
• (“focal liver lesions” OR “FLLs” OR “hepatic focal lesions” OR “liver tumor” OR “hepatic tumor”)
• AND (“artificial intelligence” OR “machine learning” OR “neural networks” OR “deep learning” OR “automated diagnosis” OR “computed tomography” OR “CT” OR “magnetic resonance imaging” OR “MRI” OR “computer-aided diagnosis” OR “automated CT” OR “automated MRI”)
Key terms such as “Artificial Intelligence,” “Cholangiocarcinoma,” “Hepatocellular Carcinoma,” and “Liver Cancer” were incorporated to refine the search and ensure inclusion of the most relevant studies.
Eligibility Criteria
Studies were included if they met all of the following criteria:
• Publication period: Published between 2000 and 2025, with 2000 chosen to reflect the emergence and growth of AI applications in medical imaging.
• Study focus: Specifically investigated AI applications for the detection or diagnosis of hepatocellular carcinoma (HCC), cholangiocarcinoma (CCA), or liver cancer.
• Study type: Original research articles with full-text availability.
• Language: Written in English, acknowledging the potential for language bias.
• Methodological rigor: Studies were required to be methodologically sound and provide complete datasets, including sufficient data to assess AI diagnostic performance metrics (e.g., sensitivity, specificity, accuracy, and AUC).
Studies were excluded if they met any of the following conditions
• Publication period: Published outside the specified date range (2000–2025).
• Duplicates: Duplicate records identified across multiple databases.
• Study type: Reviews, case reports, conference abstracts or proceedings, letters, or editorials.
• Study focus: Investigations of cancers other than HCC or CCA, or studies including mixed cancer populations without separate analyses for HCC or CCA.
• Data availability: Studies lacking full-text access or providing incomplete, unclear, or insufficient data for extraction and evaluation of AI diagnostic performance.
Study Selection
The initial database search retrieved a total of 245 articles: 133 from PubMed, 65 from Scopus, 40 from ScienceDirect, and 7 from Google Scholar. After removing duplicates and applying the eligibility criteria, 236 articles remained. Title and abstract screening identified 183 potentially relevant studies, and subsequent full-text assessment resulted in 44 studies that met all inclusion criteria.
The selection process is summarized in a PRISMA flow diagram (Figure 1).
Data Extraction and Synthesis
Data were independently extracted by three reviewers using a standardized form to ensure consistency and minimize bias. Extracted variables included:
• Author names and year of publication
• Study location and population characteristics
• Sample size and study design
• Imaging modalities employed (CT, MRI)
• AI methodologies applied (deep learning, machine learning, hybrid approaches)
• Diagnostic performance metrics (sensitivity, specificity, accuracy, AUC)
• Validation strategies (e.g., k-fold cross-validation, external datasets)
• Risk of bias assessment using the QUADAS-2 tool, evaluating patient selection, index test, reference standard, and flow and timing.
Extracted data were organized into structured tables to facilitate comparisons. AI techniques were evaluated for diagnostic capabilities in differentiating HCC and CCA [6, 8, 11].
Quality Assessment
The methodological quality of the included studies was evaluated using the QUADAS-2 tool, which assesses risk of bias across four domains: patient selection, index test, reference standard, and flow/timing. For this review, additional emphasis was placed on:
• Relevance and appropriateness of the AI model for liver cancer diagnosis.
• Adequacy of sample size to ensure statistical reliability.
• Clarity and transparency of the AI methodology, including data preprocessing, feature extraction, and training-validation strategies.
• Accuracy, completeness, and reporting transparency of diagnostic metrics.
Further considerations included the imaging modalities applied (CT and MRI) and the potential impact of language bias, as only English-language publications were included.
Results Interpretation
Extracted data were synthesized thematically and analyzed according to diagnostic performance, processing efficiency, specificity, and AI model architecture. Studies consistently demonstrated that AI models achieved diagnostic accuracy comparable to, and in some cases exceeding, that of experienced radiologists. Reported accuracy rates in differentiating HCC from CCA frequently approached or exceeded 90%, with sensitivities and specificities in the 80–95% range [8, 11].
Performance was further enhanced in studies integrating multimodal imaging or radiomic–genomic fusion approaches.
This rigorous methodology ensured a transparent and high-quality systematic review, providing critical insights into technological trends, clinical applicability, and the practical utility of AI-driven liver cancer diagnosis.
Results
Study Characteristics and Imaging Modalities
A total of 44 studies were included in this systematic review, all focusing on the application of artificial intelligence (AI) in liver cancer, particularly differentiating hepatocellular carcinoma (HCC) from intrahepatic cholangiocarcinoma (ICC) or other liver malignancies (Table 1). The studies varied in design, sample size, and imaging modality but consistently evaluated AI-based approaches particularly deep learning and radiomics models for improving diagnostic accuracy.
Retrospective vs. Prospective Designs
Most studies adopted a retrospective design, leveraging pre-existing datasets to train and test AI models. Prospective studies were less common but provided opportunities for real-time patient data collection and external validation, enhancing clinical applicability.
Imaging Modalities
Contrast-enhanced CT and MRI were the most frequently utilized imaging modalities. CT provided detailed assessment of tumor vascularity, essential for differentiating HCC from ICC. MRI offered superior soft-tissue contrast, particularly when employing dynamic contrast-enhanced (DCE) and diffusion-weighted imaging (DWI) sequences.
A limited number of studies explored hybrid modalities such as PET/CT, whereas ultrasound was rarely applied and demonstrated comparatively lower diagnostic yield.
AI Models and Approaches
• Model Architectures: Convolutional Neural Networks (CNNs) were predominant, capitalizing on automated feature extraction from raw imaging data. More advanced architectures, including Residual Networks (ResNets) and Fully Convolutional Networks (FCNs), were applied for improved segmentation and classification. In select studies, Recurrent Neural Networks (RNNs) and Long Short-Term Memory (LSTM) models were incorporated for analyzing temporal imaging sequences.
• Input Modalities and Data Types: Most models were trained on CT or MRI datasets, with some integrating advanced MRI sequences. Studies combining radiomic features (e.g., texture, shape, and intensity metrics) with deep learning demonstrated superior performance. Multi-modal approaches integrating CT and MRI achieved the highest accuracy by leveraging complementary anatomical and functional information.
Diagnostic Accuracy and Performance Metrics
AI models reported sensitivity and specificity between 75–100%, with overall accuracy ranging from 73% to 96%. The area under the curve (AUC) consistently fell between 0.74 and 0.99, indicating strong diagnostic capability. Models integrating multi-modal imaging or radiomic–clinical features often outperformed single-modality approaches.
Figure 2 illustrates the reported AUC ranges across studies. The majority clustered between 0.75 and 0.90, with a subset achieving near-perfect performance (AUC >0.95).
Figure 3 summarizes the diagnostic performance across imaging modalities. MRI and CT demonstrated the highest mean AUC values, particularly when integrated with radiomic or clinical features. Hybrid approaches (multi-modal fusion) consistently outperformed single-modality models, whereas clinical-only datasets yielded lower performance (AUC ~0.70).
Subgroup and Stratified Analyses
Tumor characteristics (size, location, microvascular invasion) and patient demographics influenced diagnostic performance. MRI-based models showed higher sensitivity for small lesions, whereas CT-based models were more effective for large tumors or vascular involvement. Multi-modal integration consistently enhanced performance.
Risk of Bias
Most studies were rated as having low-to-moderate risk of bias using QUADAS-2, with limitations primarily due to retrospective designs and limited external validation. Independent validation cohorts were relatively few, affecting generalizability.
Key Takeaways
1. AI models, particularly CNN-based and radiomics–integrated approaches, achieved high diagnostic accuracy, frequently matching or exceeding experienced radiologists.
2. MRI-based models were more sensitive for small lesions, highlighting the importance of imaging modality selection.
3. Multi-modal integration of CT, MRI, and radiomic features enhanced overall diagnostic performance.
4. Widespread clinical adoption requires large-scale, multi-center validation with standardized imaging protocols.
Discussion
This systematic review provides a comprehensive synthesis of artificial intelligence (AI) applications in liver cancer, particularly in differentiating hepatocellular carcinoma (HCC) from intrahepatic cholangiocarcinoma (ICC) and other liver malignancies. Across 44 included studies, AI models demonstrated strong diagnostic performance, with sensitivity and specificity generally ranging from 75–100%, overall accuracy between 73–96%, and area under the curve (AUC) values spanning 0.74 to 0.99. These findings highlight the rapid evolution of AI tools, which increasingly match or even surpass the performance of expert radiologists in complex diagnostic tasks [2, 5, 6, 12, 17].
Imaging modalities and diagnostic implications
CT and MRI emerged as the dominant imaging modalities applied in AI studies. CT provided robust assessment of tumor vascularity, which is essential in differentiating HCC from ICC [2, 5]. By contrast, MRI—especially dynamic contrast-enhanced (DCE) and diffusion-weighted imaging (DWI) sequences-proved more sensitive for small lesions and intratumoral heterogeneity [8, 9, 18].
These modality-specific strengths translated into differential diagnostic advantages: MRI-based models were superior in early tumor detection, whereas CT-based models were more effective for evaluating vascular invasion and larger lesions. Hybrid modalities, including PET/CT, and ultrasound-based AI approaches were less common but demonstrated potential when combined with radiomics or machine learning techniques [2, 12]. This suggests that multimodal fusion strategies may represent the most effective approach for maximizing diagnostic accuracy.
Importantly, imaging access differs by region. In Southeast Asia, where ICC prevalence is disproportionately high due to liver fluke infection, CT is more widely available than MRI [14–16, 19]. Thus, tailoring AI models to available imaging infrastructure may improve diagnostic equity in resource-limited regions.
AI model architectures and feature integration
Methodologically, convolutional neural networks (CNNs) were predominant, with newer studies employing residual networks (ResNets), fully convolutional networks (FCNs), and deep transfer learning architectures [7, 13]. These advances improved tumor segmentation and classification, even in atypical imaging presentations [5, 12]. Radiomics capturing quantitative features such as texture, shape, and intensity was widely integrated with clinical data to form fused models [12, 20].
Evidence consistently showed that combined radiomics-clinical models outperformed single-domain approaches. For example, multi-center studies employing clinical, radiomic, and genomic signatures reported significantly enhanced performance in predicting recurrence and microvascular invasion (MVI). Multi-omics frameworks further expanded predictive capability: transcriptomic signatures [21] and hepatitis B virus (HBV)-specific risk biomarkers [22] have been successfully integrated into AI pipelines, demonstrating how computational models can inform both diagnosis and therapeutic stratification.
These findings underscore that AI’s clinical utility extends beyond image interpretation toward precision hepatology and personalized treatment decision-making [12, 17].
Diagnostic accuracy and subgroup performance
The pooled diagnostic performance across studies was encouraging, with most models achieving AUCs between 0.75 and 0.90, and several exceeding 0.95. MRI-based models offered superior sensitivity for small lesions [9,18], whereas CT-based approaches proved advantageous for evaluating macrovascular invasion and resectability [2, 5]. Multi-modal fusion consistently outperformed single-modality approaches [12, 18].
Subgroup analyses revealed that tumor characteristics (size, location, vascular involvement) and patient demographics influenced AI model performance. For instance, younger patient cohorts and smaller tumor subsets were more accurately assessed with MRI-based approaches, while CT was more reliable for large and vascularly complex lesions. Such stratified analyses indicate that modality selection should be tailored not only to tumor biology but also to patient demographics.
Sources of bias and methodological limitations
Despite promising findings, several methodological limitations warrant attention. The majority of studies were retrospective, relying on pre-existing datasets [1, 5]. Prospective and multicenter validation studies remain scarce, which limits generalizability.
Imaging heterogeneity arising from differences in acquisition protocols, scanner hardware, and contrast administration posed additional barriers to reproducibility. These concerns highlight the need for harmonization of radiomic features and standardized imaging protocols [20]. High-dimensional data and relatively small sample sizes also raise risks of overfitting, while class imbalance in datasets may distort model performance [4, 13].
Another critical challenge is explainability: many deep learning models function as “black boxes,” making clinical adoption more difficult in high-stakes decision-making contexts [4, 13].
Clinical implications and public health perspectives
AI tools have the potential to substantially impact hepatology practice. Integration into clinical workflows could reduce inter-observer variability, enhance diagnostic efficiency, and provide decision support [10, 12, 18]. Comparative analyses suggest that AI systems often complement rather than replace radiologists, improving detection and classification when used in tandem [2, 5, 10].
Beyond diagnosis, AI has shown value in predicting recurrence, treatment response, and MVI, with implications for surgical planning, transplant eligibility, and allocation of therapies such as TACE and immunotherapy [9, 23]. These prognostic utilities are particularly important in high-burden regions, where liver cancer outcomes remain poor and resources are constrained [14–16, 19]. Cost-effectiveness analyses and implementation studies will be critical in bridging the gap between technical performance and real-world utility [4].
Future directions
To advance clinical adoption, future research should prioritize large-scale, prospective, multicenter studies with external validation [1, 20]. Standardization of imaging acquisition and radiomic feature extraction is urgently needed to ensure reproducibility across institutions [20]. Multi-omics integration linking imaging, clinical, and genomic features represents a frontier for precision hepatology [21, 22].
Development of explainable AI systems will be essential to enhance clinician trust and regulatory approval [4],130. Moreover, AI should be designed for seamless integration into existing radiology and oncology workflows, with attention to user interface, interpretability, and clinical decision support. Finally, consideration of public health perspectives particularly in high-burden regions such as Asia—will be vital to ensuring that AI technologies address global disparities in liver cancer care [14–16, 19].
In conclusion, AI has demonstrated substantial promise in differentiating HCC from ICC, with performance metrics frequently rivaling expert radiologists. While challenges remain in terms of reproducibility, standardization, and clinical translation, ongoing advancements in multi-modal integration, multi-omics approaches, and explainable AI systems point toward a future where AI can play a central role in precision hepatology and global liver cancer control.
Author Contribution Statement
All authors contributed equally in this study.