Primary liver cancer is the sixth most commonly diagnosed cancer worldwide and hepatocellular carcinoma (HCC) accounts for over 80% of newly diagnosed cases
[1]. Due to the lack of early diagnostic markers, absence of specific symptoms in early disease, and inadequate screening programs in most countries, nearly two out of every three patients diagnosed with HCC have intermediate or advanced disease upon diagnosis
[2][3]. Unfortunately, these patients often have a poor prognosis as their disease may not be amenable to curative interventions, such as surgical resection, liver transplantation, or ablation, and, until recently, effective systemic options were limited. As a result, HCC is the fourth most common cause of cancer-related death worldwide with a relative five-year survival rate of 18%
[4][5]. However, the recent inclusion of immunotherapy into the HCC treatment paradigm and the expansion of downstaging/bridging protocols to liver transplantation have improved overall survival and established a new standard of care for patients with HCC
[6][7]. In addition to surgical and systemic therapeutic options, a number of liver-directed therapies (e.g., bland transarterial embolization (TAE), transarterial chemoembolization (TACE), radioembolization (RE)) are available depending on the extent of cirrhosis and the stage of the disease
[5]. With the expanding therapeutic armamentarium for patients with HCC, novel tools are needed to effectively stratify patients to maximize therapeutic benefit.
Artificial intelligence (AI) has recently emerged as a viable clinical tool with growing utility in the management of HCC. Broadly, AI is a subdivision of data science which describes the theory and development of computer systems that perform tasks requiring human-level intelligence such as visual perception or decision making. AI was originally conceptualized in the 1950s by the mathematician Alan Turing and the field has greatly expanded since its original conception. As technology continues to evolve, new AI techniques have been developed to address more complex and sophisticated problems. Machine learning and deep learning are two such subfields of AI (Figure 1).
2. Screening and Detection
Since many patients diagnosed with HCC have advanced disease and limited therapeutic options, improvements in screening and early detection are necessary to improve outcomes. For example, a review of country-level HCC surveillance programs demonstrated that countries with established and effective HCC surveillance programs detected HCC at significantly earlier stages and had lower overall mortality
[15]. Effective programs identify and stratify patients at high-risk of developing HCC and enroll them into regular surveillance protocols such as biannual evaluation with liver ultrasound and liquid tumor markers. Ideally, healthcare professionals intervene prior to the development of irreversible HCC risk factors (i.e., cirrhosis). However, cost/benefit restraints and high false-positive rates limit the utility of screening low-prevalence populations. AI techniques may help us to overcome these limitations (
Table 1). In a 2022 study, Blanes-Vidal et al. evaluated asymptomatic patients from a primary care population without a prior diagnosis of liver disease
[16]. The authors tested the diagnostic performance of ensemble models, a machine learning approach, to detect liver fibrosis and then compared the performance with standard blood-based scoring systems. The ensemble models included data readily available during a primary care visit. In a subset of 463 patients that received a liver biopsy, the ensemble learning models significantly outperformed standard blood-based indices to detect liver stiffness (>8 kPa) and fibrosis (Kleiner biopsy stage F2 to F4) with AUCs of 0.86–0.94 vs. 0.60–0.76. Furthermore, all the ensemble models had a ≥98% negative predictive value. Similarly, other AI models have demonstrated superiority to standard models in predicting the development of HCC
[17][18].
Radiomics, a quantitative method to extract features (e.g., shape, intensity, texture) from medical imaging, can significantly improve the diagnostic yield of imaging modalities, especially when combined with other AI techniques such as deep learning. While established criteria exist for the radiologic diagnosis of HCC in high-risk patients (i.e., LI-RADS), the true proportion of patients with LI-RADS 5 lesions (diagnostic of HCC) is unclear. Additionally, this classification system only applies to patients at high risk of HCC (e.g., cirrhosis, HBV infection)
[19]. The diagnostic uncertainty of many newly identified liver lesions may lead to unnecessary serial imaging, invasive procedures (e.g., biopsy), and undue psychological stress on the patient. For these reasons, much energy has been devoted to building radiomic-based models that can accurately detect and diagnose HCC (
Table 1). Yasaka et al. developed a deep learning model using a convolutional neural network (CNN) to differentiate malignant liver lesions from non-malignant lesions
[20]. In this retrospective study, the authors built a CNN using imaging sets from 460 patients who were found to have liver lesions on triple-phase CT. The lesions were classified into five categories as follows: classic HCC (A), malignant tumors other than HCC (B), indeterminate masses and rare benign liver masses (C), hemangiomas (D), or cysts €. The CNN accurately identified malignant lesions with the median AUC of 0.92 for differentiating categories A–B from categories C–E. Furthermore, Mokrane et al. used a radiomic model to improve the diagnostic accuracy of HCC in cirrhotic patients with indeterminate liver nodules
[21]. In this multicenter retrospective study, multiphasic CT scans of 178 cirrhotic patients who had undergone the biopsy of indeterminate liver nodules were included. Nearly 14,000 quantitative features were extracted from imaging sets. With machine learning algorithms, a radiomic signature was created and validated to classify these indeterminate liver nodules as HCC or non-HCC. While the LI-RADS scores were similar between the HCC and non-HCC groups, the radiomic signature reached an AUC of 0.70 in the discovery cohort and 0.66 in the validation cohort. AI has also been used to augment the diagnostic capabilities of other imaging modalities including ultrasound and magnetic resonance imaging (MRI)
[22][23][24].
While the generalizability of screening and diagnostic models derived from AI techniques requires further evaluation, these models offer non-invasive and resource-efficient means to reliably screen and detect HCC in select populations. This technology can be especially useful in low resource centers where specialized liver radiologists may not be readily available to aid in the diagnosis of complex liver lesions. Furthermore, these models may help institutions and governments to more efficiently allocate scarce resources to HCC screening and intervention due to improved risk stratification. For example, using a deep learning recurrent neural network model to generate HCC risk scores, Ioannou et al. found that 80% of HCC cases diagnosed in the subsequent three years occurred in the highest 51% of risk scores
[18]. Risk-based screening supported by AI algorithms may increase diagnostic yield, optimize resource utilization, and help us to overcome the suboptimal performance of existing tools
[25].
Table 1. Select studies utilizing AI in screening and diagnosis of HCC.
3. Prognosis and Treatment
3.1. HCC Prognosis and Risk of Recurrence
Recent multi-omic investigations have advanced the understanding of the carcinogenic mechanisms responsible for HCC, revealing many potential biomarkers
[26]. Evaluating these large data sets with AI methods may improve current prognostic ability by identifying more aggressive subtypes and patients at high risk of recurrence (
Table 2). In an early study using a deep learning framework to integrate multi-omic data in patients with HCC, Chaudhary et al. developed a deep learning model from 360 patients with HCC using RNA sequencing, miRNA sequencing, and methylation data from The Cancer Genome Atlas
[27]. The model identified a more aggressive subtype with worse survival characterized through frequent TP53 mutations; a higher expression of KRT19, EPCAM, and BIRC5; and activated Wnt and Akt signaling pathways. The investigators then validated the model on five external data sets with acceptable results.
In addition to detecting relationships within large multi-omic data sets to improve prognostication, AI techniques can help us identify biomarkers in the preoperative setting typically only identified through pathologic evaluation such as microvascular invasion (MVI)
[28]. For example, multiple studies have shown the feasibility of using machine learning algorithms to accurately predict the presence of MVI based on preoperative axial imaging characteristics
[29][30][31]. Chong et al. built a radiomic-based nomogram to assess the risk of MVI
[32]. This retrospective study analyzed preoperative MRIs from 356 patients with pathologically confirmed solitary HCC less than 5 cm. The nomogram, built by extracting radiomic features from images containing tumors, peritumoral tissue, and non-tumoral liver parenchyma, accurately predicted the risk of MVI as well as recurrence-free survival.
Similarly, other groups have identified biomarkers of HCC recurrence using machine learning methods (
Table 2)
[33]. Yan et al. created a deep learning MR signature derived from imaging characteristics including tumor size, arterial phase enhancement type, capsular appearance, presence of a hypointense halo, intratumoral necrosis, satellite nodules, and peritumoral hypointensity
[34]. When combined with clinical factors such as MVI and tumor number, the MR signature predicted early recurrence better than clinical data alone. Another model created by Ji et al. extracted radiomic features from preoperative CT scans to build a radiomic signature that, when integrating clinical data such as MVI, AFP level, and tumor number, accurately predicted the risk of recurrence in patients after resection of early-stage HCC
[35]. Improved prognostication and risk stratification with AI techniques may better inform management decisions for patients with HCC.
Table 2. Key studies utilizing AI to predict prognosis and risk of recurrence of HCC.
3.2. Pathologic Assessment
AI techniques have also been applied to evaluate associations between histologic features and outcomes in numerous disease processes, including HCC (
Table 3)
[36][37][38][39]. Using whole-slide imaging, Yamashita et al. developed a deep learning-based system to predict a recurrence-free disease interval after curative-intent hepatectomy in patients with HCC. Their model stratified patients into low-risk and high-risk subgroups and outperformed the standard tumor–node–metastasis (TNM) staging system
[40].
Chen et al. built a neural network that was able to assist in the prognostication of HCC based on histologic whole-slide imaging. The model used hematoxylin and eosin slides from a genomic database to train a neural network to classify liver lesions as malignant with 96.0% accuracy and predict lesion histopathological grade with 89.6% accuracy. Furthermore, the model also predicted select gene mutations including CTNNB1, FMN2, TP53, and XFZ4
[41]. AI-based pathology models can also predict the activation of immune signatures in HCC. Zeng et al. used deep learning approaches on whole-slide histologic images and gene expression profiling, derived from the Cancer Genome Atlas Liver Hepatocellular Carcinoma (TCGA-LIHC) public dataset, to develop models that predicted the activation of six key immune signatures that, when overexpressed, correlated with the response to immunotherapy
[42]. While prospective validation studies are needed, these data demonstrate the potential utility of AI to select patients who will have a greater response to immunotherapy and may inform adjuvant therapy decisions.
Table 3. Key studies demonstrating the use of AI on whole-slide imaging.
3.3. Locoregional Therapies
AI-based models may help to select patients who are good candidates for locoregional therapies such as radiofrequency ablation (RFA) (
Table 4). Wu et al. built an artificial neural network based on fifteen clinical variables from HCC patients who had undergone CT-guided RFA
[43]. Variables included patient characteristics, tumor size, tumor number, and laboratory values (e.g., AFP). Ultimately, the artificial neural network model predicted one-year disease-free survival with an AUC of 0.84 and one-year disease-free survival with an AUC of 0.75. Lui et al. created a deep learning radiomic-based model of preoperative contrast-enhanced liver ultrasound images and predicted the response to the first TACE session in patients with HCC
[44]. In another study, investigators trained and validated a CNN to automatically assess splenic volume. Patients with higher splenic volumes, as automatically determined by the CNN, had significantly higher risk of hepatic decompensation and lower overall survival after TACE. Conversely, axial and craniocaudal splenic diameter did not correlate with outcomes
[45].
3.4. Automatic Methods for Liver and Tumor Segmentation
Automatic liver and tumor segmentation has numerous clinical applications which may aid in the management of HCC such as the preoperative volumetric quantification and assessment of treatment responses to locoregional therapies (
Table 4). The use of manual and semi-automatic liver segmentation methods can lead to inconsistencies due to user variability. Furthermore, segmentation is often a slow, labor-intensive process. With advancements in deep learning techniques and CNNs, completely automatic liver segmentation is now feasible
[46][47][48]. However, automatic algorithms do have their own limitations including the need for large training datasets to develop accurate algorithms. Automatic tumor segmentation can be even more challenging as tumor characteristics, such as variability in size, variability in location, and indiscrete borders between healthy liver parenchyma and tumors, can decrease the accuracy of models. Regardless, early studies demonstrate that automatic segmentations outperform semi-automatic methods with regard to the accuracy and repeatability of segmentation
[49]. Meng et al. used a three-dimensional (3D) dual-path multiscale CNN to build a liver and liver tumor segmentation algorithm based on abdominal CT images. The dual path multiscale 3D network architecture allowed the model to capture large scale global features through one path while capturing more granular local features through the second path. Notably, their algorithm performed best on large tumors
[50]. Zheng et al. built a four-dimensional (4D) deep learning model to segment HCC lesions based on dynamic contrast-enhanced MRIs. In an attempt to avoid the “black box” learning that occurs with many automatic segmentation models, the investigators first build a 3D CNN to separately extract imaging features from each individual phase of imaging. From there, these data were fed into a convolutional long short-term memory (C-LSTM) network module in order to extract higher-level features, including temporal information and dynamic features, that varied through the multi-phase imaging. This allows us to use dynamic features that are not only characteristic of HCC but critical to diagnosis, such as arterial phase hyperenhancement and portal venous or delayed-phase washout. Their model achieved a Dice score of 0.825 ± 0.077 for HCC tumor segmentation in an internal test set and 0.786 ± 0.073 in an external set, implying good generalizability of this model
[51].
3.5. Surgical Complications
Many HCC patients have some level of hepatic dysfunction upon presentation, increasing the risk of post-hepatectomy liver failure (PHLF). PHLF remains the most frequent cause of postoperative mortality in patients following hepatectomy, highlighting the importance of identifying patients at high risk of PHLF prior to resection to more effectively select patients for adjuncts such as portal vein embolization. Wang et al. constructed a machine learning clinical model using laboratory values, tumor characteristics, and surgical variables (e.g., surgical approach, extent of resection, intraoperative blood loss) to predict the risk of PHLF. The model outperformed traditional models such as MELD, Child-Turcotte-Pugh, or albumin-bilirubin grade when predicting PHLF
[52]. AI-derived algorithms have successfully predicted other surgical complications. Laino et al. predicted the risk of a postoperative bile leak following hepatectomy using a combined clinical–radiomic model in 378 patients with a preoperative CT. Radiomic data was extracted from a virtual liver biopsy obtained from a 2 cm cylinder of non-tumoral liver parenchyma on the portal phase of preoperative CT. Their combined model (AUC 0.74) outperformed a model using clinical data alone (AUC 0.61) to predict the risk of a bile leak
[53].
Table 4. Key studies demonstrating the use of AI in management of HCC.
4. Intraoperative Use of Artificial Intelligence
A thorough understanding of the vascular and biliary anatomy from adequate multiphasic abdominal imaging is necessary to perform a safe hepatectomy and minimize complications. Three-dimensional liver reconstruction technology may improve perioperative outcomes in patients undergoing a major hepatectomy by further elucidating spatial relationships between the tumor and critical vascular and biliary structures. A meta-analysis evaluating the efficacy and safety of 3D-reconstruction liver models in patients undergoing a major hepatectomy showed shorter operative times, less intraoperative blood loss, fewer hepatic inflow occlusion events, shorter hospital stays, and fewer postoperative complications when using such technology
[54]. Deep learning algorithms can help us to automate the reconstruction of 3D liver models with reliable accuracy and detail
[55].
As an adjunct to intraoperative liver ultrasound (IOUS), machine learning algorithms may help us to overcome some limitations of traditional ultrasounds and improve the accuracy of identifying liver lesions intraoperatively. Barash et al. trained a CNN on intraoperative ultrasound imaging to detect liver lesions. The algorithm achieved an AUC of 80.2% and an overall classification accuracy of 74.6%
[56]. Furthermore, Takamoto et al. used real-time virtual sonography (RVS), an AI-assisted platform that merges preoperative CT images with real-time IOUS, to enhance IOUS with the identification of small intrahepatic lesions
[57]. The median liver lesion size was 6.0 mm and RVS significantly improved surgeon confidence in lesion identification. Importantly, of the 17 lesions undetectable using fundamental IOUS, 14 were identified through RVS and ultimately treated.
However, despite the successful applications of AI technology during liver surgery, several challenges remain. This is especially true of 3D overlays, which are difficult to use during hepatectomy. The mobilization of the liver and the deformation of the parenchyma during hepatectomy make real-time overlays onto tissues more challenging and less accurate. This is in comparison with other surgical disciplines such as neurosurgery, where the target—in this case the brain—is rigid and fixed, allowing for an easy overlay of 3D reconstructions. Preliminary studies demonstrate the feasibility of a physics-based elastic augmented reality model that can provide a real-time 3D overlay during hepatectomy, allowing for the deformation and mobilization of the tissue; however, further quality improvement needs to occur prior to meaningful use during hepatectomy
[58].