Primary liver cancer is a challenging disease that was the second most common cause of cancer mortality globally in 2018, and it is the 7th most common type of cancer (Bray et al., 2018). Between 2020 and 2040, the number of new cases of liver cancer are predicted to increase by 55% per year (Rumgay et al., 2022). Liver cancer cells are differentiated into primary and secondary liver cancers according to the origin of the cancer cells. In primary liver cancer, the cancer originates within the liver itself, while secondary liver cancer is a result of the metastasis of other organs.

Primary liver cancer types include hepatocellular carcinoma (HCC), intrahepatic cholangiocarcinoma, hepatoblastoma, fibrolamellar carcinoma, angiosarcoma, and hemangiosarcoma (Astrologo et al., 2023). HCC accounts for approximately 75% of all liver cancer cases. Infection with hepatitis B and C is the major risk factor for HCC; however, other factors may play a role, such as aflatoxin exposure, alcohol consumption, and smoking (Chuang et al., 2009; Mittal and El-Serag, 2013). The treatment of hepatocellular carcinoma depends on several factors, including tumor size, cancer stage, extrahepatic metastasis, and the extent of vascular invasion. In general, patients with HCC have a poor prognosis, which is determined by the stage of liver disease, disease severity, and diagnosis timing. Therefore, early diagnosis is crucial for a better prognosis (Galle et al., 2018; Marrero et al., 2018).

Serum biomarkers, such as alpha-fetoprotein (AFP-L3), des-gamma-carboxy prothrombin (DCP), Golgi protein 73 (GP73), and glypican-3 (GPC3), have beneficial value for early diagnosis. Several trace chemicals, such as circulating tumor noncoding RNA (ct-ncRNA), cell-free DNA (cfDNA), circulating tumor DNA (ctDNA), and circulating tumor cells (CTCs), are released into biological fluids and could serve as valuable diagnostic agents (Di Tommaso et al., 2009; Toyoda et al., 2011; Guo et al., 2018; Choi et al., 2019). Fine-needle aspiration (FNA) biopsy is considered an additional confirmation test (Galle et al., 2018). Computed tomography (CT) and magnetic resonance imaging (MRI) are widely used in cancer monitoring and diagnosis. Nevertheless, their sensitivity and specificity for early HCC detection are relatively low, so liver-specific contrast agents are used to improve imaging accuracy. The combination of gadolinium-ethoxybenzyl diethylenetriamine pentaacetic acid (Gd-EOB-DTPA) and MRI improves the diagnosis of liver lesions, but this combination is not optimal for small lesion detection (Yu et al., 2014; Marrero et al., 2018). Furthermore, imaging techniques depend on observer interpretation, which represents a major source of error and misdiagnosis; therefore, artificial intelligence (AI)-based models have been developed to overcome this issue.

Radiomics analysis is a novel tool developed for extracting data from medical images and combined with imaging techniques for better performance (Lambin et al., 2012; Wang et al., 2023). Other AI models are based on machine learning, deep learning (DL), and convolutional neural networks (CNNs). The algorithms generated by machine learning must first undergo training on datasets to make predictions. Deep learning is a subset of machine learning that learns and extracts difficult data using multiple layers. Another technology is convolutional neural networks (CNNs), which are considered the ideal model for diagnosis because they can process complex visual data through multiple layers and filters. Different models have been developed in conjunction with traditional methods to optimize the diagnosis process. AI tools are unbiased, smart, cost-effective, and noninvasive, and their efficacy is comparable to that of humans (Yamashita et al., 2018; Saba et al., 2019; Awal et al., 2023). In this systematic review and meta-analysis, we aimed to evaluate the diagnostic performance of different AI models for the diagnosis of hepatocellular carcinoma in comparison with human expertise.

First of all, it is necessary to outline and explain the parameters of common effect and random effects models employed in the analysis. The fixed-effect model, also known as common-effect model assumes one true effect size underlying all studies included in this meta-analysis. In cases where studies are assumed to be similar enough in design and population, that’s when this model is appropriate. The key parameter for this particular model is pooled effect size which represents average of the sizes of the effects obtained from all the studies.

Contrarily, random-effects model believes that there is a difference in true effect sizes among studies. It takes into consideration variability within and between variables across various research thus becoming an appropriate choice if considerable heterogeneity exists amongst several researches. Within this specific model the vital parameters include pooled effect size and between-study variance commonly referred to as Tau² indicating how much there exists variation amidst true effects.

The between-study variance in a meta-analysis is calculated by Tau² (Tau-squared) and it helps us understand how much the effect sizes vary across studies over and above what would be expected by chance alone. Although there is no exact cutoff, we can interpret small heterogeneity of values near 0, moderate heterogeneity ranging from 0.01 to 0.1, and large heterogeneity indicated by values greater than 0.1 (West et al., 2010). For example, in this context, a Tau² value of 0.0057 indicates that there is moderate heterogeneity which implies that while there may be some variation of the effect size found across the studies; it is not excessive enough to be called very high heterogeneity as such. In other words, differences between results of these studies are due to more than random occurrences but are not too far apart.

In this systematic review and meta-analysis, we aimed to explore the potential of artificial intelligence tools for the diagnosis of hepatocellular carcinoma compared with human expertise. New models are being developed daily, which is considered an invaluable opportunity for advancing diagnostic accuracy and saving doctors’ time. Several scientists have debated whether these tools can replace humans in the future. Our study is the first systematic review to compare AI-based tools with physicians in the diagnosis of HCC. Early diagnosis of HCC is pivotal because patient survival is linked to hepatocellular carcinoma staging; patients diagnosed in early stages have higher five-year survival rates than those diagnosed in late stages (70 and 20%, respectively). Therefore, there is an urgent need to develop accurate and sensitive tools to optimize the diagnostic process for HCC (American Cancer Society, 2020).

Our analysis showed that most AI models are more sensitive than physicians, except for Kim et al. (2020), in which the expert radiologist had greater sensitivity than the AI model (98 vs. 87, respectively); however, they had almost the same specificity (93%). The proposed model had a sensitivity similar to that of a nonexpert radiologist (approximately 86.5%) (Kim et al., 2020). In some cases, the diagnosis of HCC using imaging techniques is quite challenging and requires strong experience, as there is great heterogeneity within HCC cells; different areas can have different growth patterns and levels of differentiation (Quaglia, 2018).

To address this issue, Zhen et al. (2020) developed an AI model that combines MR images and clinical data. This combination significantly improved the classification ability of the new models [AUC = 0.985 (95% CI, 0.960–1)]. Furthermore, Zhen et al. (2020) investigated the impact of sequence number on model performance. Different imaging sequences were used to train the model. Six sequences (contrast-enhanced T1 sequence, contrast-enhanced T1 sequence, diffusion, late arterial, portal venous, and equilibrium) or three sequences (T1, T2, and diffusion) as well as clinical data were utilized as model inputs. When clinical data were built into the model, the number of sequences did not have an impact on the performance. However, for models based on images only, model A, which had only three unenhanced sequences (AUC = 0.925, 95% CI = 0.871, 0.978), had better results than model B, which had 0.879, 95% CI = 0.813, 0.9452 (Zhen et al., 2020). Gao et al. (2021) measured the sensitivity and accuracy of AI-assisted physicians versus the AI model (STIC) or physicians alone. The AI-assisted physicians and the STIC model had approximately comparable sensitivities of 83.3 and 86.5%, respectively. Physicians had the lowest sensitivity (78.4%), but their accuracy did not significantly differ from that of the STIC model (70.8 and 72.6%, respectively) (Hamm et al., 2019; Kim et al., 2020; Zhen et al., 2020; Gao et al., 2021; Nishida et al., 2022; Liu et al., 2023; Urhuț et al., 2023).

The most sensitive model was the faster region-based convolutional neural network (RCNN) proposed by Liu et al. (2023). Models based on RCNNs exhibit a good ability to extract data from various images (Hamm et al., 2019; Kim et al., 2020; Zhen et al., 2020; Gao et al., 2021; Nishida et al., 2022; Liu et al., 2023; Urhuț et al., 2023). Raimundo et al. (2023) developed an AI model that combines RCCN and MRI for the diagnosis of breast cancer. Similarly, this model achieved a high accuracy of 94.4%. Another model by Liu et al. (2022) was developed based on the RCNN and medical images for diagnosing bile duct tumor thrombi in patients with HCC. In addition, this model showed a high sensitivity of 94% and a good specificity of 78%, which confirms the efficiency of RCNN-based models in improving the diagnosis of medical images (Liu et al., 2022, 2023; Raimundo et al., 2023).

Several diagnostic AI models based on convolutional neural networks (CNNs) have evolved because they do not require a clear definition of the lesion to interpret the images. Yamashita et al. (2020) proposed a CNN-based model that could classify hepatic observations without predefining hand-crafted imaging features with 60.4% accuracy. Similarly, Wang et al. (2019) investigated a convolutional neural network (CNN)-based model for the differentiation of liver masses via dynamic contrast agent-enhanced computed tomography (CT). The model achieved a median diagnostic accuracy of 0.84 (Yasaka et al., 2018).

Nishida et al. (2022) developed three different models to establish the relationship between the training dataset size and CNN performance. Model 3, which had the most training data, performed better than the other two models and the physicians did. Increasing the training set size has a proven impact on improving the accuracy of CNN-based models; nevertheless, large training sets consume more money, time, and effort (Radiuk, 2017; Nishida et al., 2022).

Clinical data have a major influence on imaging interpretation. Urhuț et al. (2023) compared the performance of novel AI-based contrast-enhanced ultrasound (CEUS) with that of two groups of clinicians: the first group knew all relevant clinical data, while the second group was blinded. The sensitivity and accuracy of the blinded group were lower than those of the unblinded group (58.3 and 83% vs. 50 and 79.6%, respectively). Furthermore, the AI-based contrast-enhanced ultrasound (CEUS) model had high sensitivity (86.9%) but less accuracy (69.9%) than did clinicians who were aware of the patient’s case (Urhuț et al., 2023). The sensitivity of contrast-enhanced ultrasound varies across different studies. A meta-analysis involving studies conducted in 1996–2016 showed that the sensitivity of CEUS for detecting HCC was 85%, although Jiang et al. (2021) reported a lower sensitivity of 69% for lesions <20 mm and 75% for lesions ≥20 mm. The combination of CEUS and AFP levels showed promising results, and the sensitivity increased to 83.1%; therefore, there is a need to establish evidence of the efficiency of CEUS in the diagnosis of HCC and to investigate the performance of the triple combination of CEUS, AFP levels, and AI models (Zhang et al., 2017; Jiang et al., 2021). In addition to diagnosing HCC, machine learning has a promising role in liver transplantation. Machine learning-based models could provide excellent predictions of the risk of complications and short-and long-term mortality. Furthermore, they can predict posttransplant outcomes better than traditional scoring systems. The incorporation of AI into the transplantation process saves time and money and increases the rate of transportation success (Chongo and Soldera, 2024). In general, AI has great potential in diagnosing different gastrointestinal tract pathologies due to its ability to use complex mathematical data involving multiple parameters and sophisticated formulas to draw conclusions that would be challenging or unfeasible for humans to process alone. However, many concerns are associated with the potential of AI in prognostication, such as the need for high-quality images for model training, ethical considerations of data usage, and the cost–benefit ratio (Do and Gastrointestinal, 2023).

Compared with physicians, AI models have the ability to improve the diagnosis of hepatocellular carcinoma. Faster R-CNN-based models are excellent for imaging classification and data extraction. CNN-based models have high sensitivity, and increasing the size of the training database significantly augments the accuracy of CNN models. The sensitivity of CEUS in the diagnosis of HCC is debatable; however, the combination of CEUS and AI models yields high sensitivity. Despite the promising results, AI models should not entirely replace humans in the diagnostic process; rather, they should be used as an assistant tool for more accurate and less timed decisions. We need to conduct more studies on the performance of AI-assisted physicians versus physicians without assistance, considering physicians’ level of experience.

FA-O: Conceptualization, Data curation, Formal analysis, Funding acquisition, Investigation, Methodology, Project administration, Resources, Software, Supervision, Validation, Visualization, Writing – original draft, Writing – review & editing. WH: Conceptualization, Data curation, Formal analysis, Funding acquisition, Investigation, Methodology, Project administration, Resources, Software, Supervision, Validation, Visualization, Writing – original draft, Writing – review & editing. MG: Conceptualization, Data curation, Formal analysis, Funding acquisition, Investigation, Methodology, Project administration, Resources, Software, Supervision, Validation, Visualization, Writing – original draft, Writing – review & editing. NA: Conceptualization, Data curation, Formal analysis, Funding acquisition, Investigation, Methodology, Project administration, Resources, Software, Supervision, Validation, Visualization, Writing – original draft, Writing – review & editing. MA: Conceptualization, Data curation, Formal analysis, Funding acquisition, Investigation, Methodology, Project administration, Resources, Software, Supervision, Validation, Visualization, Writing – original draft, Writing – review & editing. AY: Data curation, Formal analysis, Validation, Software, Writing – review and editing. AR: Data curation, Formal analysis, Methodology, Validation, Software, Writing – review and editing.