HCC usually, though not always, develops in a cirrhotic liver. Accordingly, clinical practice guidelines recommend regular abdominal ultrasound in patients with hepatic cirrhosis; indeed, it is considered the method of choice for screening of space-occupying lesions. Ultrasound is therefore the main tool to evaluate liver disease and detect new lesions. However, image interpretation is not always easy and may present interobserver variability.

To assess the underlying disease, Bharti et al[7] proposed an ANN model to differentiate four stages of liver disease using data obtained from ultrasound images: normal liver, chronic liver disease, cirrhosis, and HCC. The classification accuracy of the model was 96.6%[7]. Liu et al[8] designed an algorithm to classify ultrasound images. They selected the liver capsule to determine the presence of cirrhosis, even in early stages when the usual findings reported by the radiologist, such as a nodular liver outline, enlarged porta or splenomegaly, are still not obvious. Using their analysis of the morphology of the liver capsule, they were able to determine the presence or absence of cirrhosis, with an area under the curve of 0.968[8].

The human yield when characterizing a liver lesion from ultrasound images is limited. Schmauch et al[9] designed a DL system able to detect and classify space-occupying lesions in the liver as benign or malignant. After a supervised training using a database of 367 images together with the radiological reports, the resulting algorithm detected and characterized the lesions with a mean receiver operating characteristic of 0.93 and 0.916, respectively. Although the system requires validation, it could increase the diagnostic yield of ultrasound and warn of possibly malignant lesions[9].

AI has also been used in contrast-enhanced ultrasound (C-US), improving its ability to identify characteristics suggestive of cancer. Guo et al[10] demonstrated that DL applied to the behavior of liver lesions seen on C-US in three phases (arterial, portal and late) increased the accuracy, sensitivity and specificity of the study[10].

When a follow-up ultrasound shows a new liver lesion, other imaging studies are undertaken, mainly dynamic contrast-enhanced computed tomography (CT) or MRI, to obtain a more precise evaluation. The radiological behavior of liver lesions in dynamic CT or MRI studies is useful for characterization of the lesion. If a liver lesion > 1 cm fulfils certain radiological criteria, almost pathognomonic for HCC such as the presence of hyperenhancement at arterial phase and washout at portal or late phases in a cirrhotic patient, no further studies are needed for its diagnosis or histological confirmation. However, liver nodules often present an indeterminate behavior on CT and a biopsy of the lesion is required, as recommended in the European Association for the Study of the Liver guidelines[11]. This, though, requires assuming the risks involved in the procedure, or close follow-up, as indicated in the American Association for the Study of Liver Diseases guidelines[12], with a high number of studies and the possibility of not diagnosing a malignant lesion in time. Mokrane et al[13] undertook a retrospective analysis of 178 patients with cirrhosis and liver nodules in whom the Liver Reporting and Data System criteria were unable to distinguish the neoplastic from the non-neoplastic lesions, thereby necessitating a biopsy; 77% proved to be malignant on biopsy. Using DL techniques to classify nodules as HCC or non-HCC achieved an area under the curve (AUC) of 0.70. Another retrospective study, by Yasaka et al[14], analyzed the yield of an ANN, composed of three layers, classifying liver masses using contrast-enhanced CT into five categories: A, classic HCC; B, malignant tumors apart from HCC (cholangiocarcinoma, hepatocholangiocarcinoma or metastasis); C, indeterminate masses, dysplastic nodules or early HCC and benign masses other than cysts or hemangiomas; D, hemangiomas; E, cysts. After supervised training using over 55000 image sets the authors obtained a high accuracy for the classification of liver lesions, especially for the differentiation between categories A-B and C-D.

Quantifying the tumor load may be useful, particularly for the detection of tumor recurrence in follow-up CT studies. As tumor relapses can be small and go unnoticed, Vivanti et al[15] described an automated detection method of recurrence, based on the initial appearance of the tumor, its CT behavior, and the quantification of the tumor load at baseline and during the follow-up. The technique had a high rate of true positives in the identification of tumor recurrence, with an accuracy of 86%.

Liver segmentation is of great importance to assess liver lesions and for planning the ideal treatment. However, manual segmentation is made more difficult by the heterogeneity of the lesions or their diffuse borders. Li et al[16] proposed a CNN that can segment liver tumors based on CT images, with an accuracy of 82.67% ± 1.43%, better than that of traditional techniques, thereby favoring suitable treatment planning.

CNN applied to MRI has also been analyzed. Hamm et at[17] developed and validated a DL system based on a CNN that classifies MRI liver lesions, with an accuracy of 92%, a sensitivity of 92% and a specificity of 98%; and an average computation time of 5.6 ms.

Other studies have associated additional MRI sequences and risk factors plus the patient’s clinical data to apply an automated classification system cataloguing liver lesions as adenoma, cyst, hemangioma, HCC and metastasis, with a sensitivity and specificity of 0.8/0.78, 0.93/0.93, 0.84/0.82, 0.73/0.56 and 0.62/0.77, respectively[18]. Zhang et al[19] described a training model using MRI in 20 patients to classify liver tissues. Their results were promising, improving the yield of the reference models used.

Preis et al[20] evaluated the yield of fluorine 18 fluorodeosyglucose positron emission tomography/computed tomography (18F-FDG PET/CT) using a neural network to analyze liver uptake of 18F together with patient and laboratory data. They achieved a high sensitivity and specificity to detect liver malignancy unidentified visually, showing that this technique could complement the radiologist in the interpretation of PET, though their main aim was to evaluate metastatic liver disease, where 18F-FDG PET/CT has greater applicability.

The histopathological classification of a liver lesion and the differentiation of the tumor strain is crucial for treatment planning and a prognostic estimate of the disease, though this can sometimes prove challenging even for expert pathologists. Kiani et al[21] used AI as a support for the pathologists, focusing on the histological differentiation between HCC and cholangiocarcinoma. They analyzed prospectively the impact of this aid in the diagnostic yield of 11 pathologists, finding that it made no change to their mean accuracy.

Others have described how a deep CNN using previous histopathological images of HCC can make an automated diagnosis of HCC and distinguish healthy tissue from tumor tissue, in addition to identifying certain biological predictors[22]. Table 1 shows the main studies using AI techniques in the diagnosis of HCC.

In usual clinical practice the diagnosis and treatment assessment of HCC are commonly done with such imaging techniques as C-US, CT and MRI, after analyzing certain tumor features like vascularization or behavior after the administration of contrast material[23]. However, these characteristics are liable to subjectivity in their interpretation by the radiologist, in addition to the lack of high resolution dimensional images. But a new technology has recently appeared in the area of radiology and cancer, radiomics[24]. Although it is not yet in extensive use in clinical practice, it has nevertheless awoken great interest. This technology allows for the extraction of a great amount of quantifiable objective data contained in the radiological images and their later association with the underlying biological processes. Analysis of all these data with AI software can provide useful diagnostic and prognostic information with predictive accuracy[24,25].

Early tumor recurrence after surgical resection is associated with a poor prognosis. The preoperative identification of patients at high risk of recurrence is fundamental to avoid unnecessary treatment. Computer models have been developed that analyze certain tumor characteristics and aid in the preoperative prediction of the risk of recurrence or the evaluation of survival after resection.

Vascular microinvasion (VMI) has been established as an independent predictive factor of recurrence, associated with poor results after tumor resection[26]. Although the preoperative availability of information about VMI would be of much benefit, radiological techniques currently in clinical use do not provide for an adequate direct diagnosis.

Several studies have managed to elaborate radiomic signatures that enable prediction of the preoperative status of VMI, based on contrast-enhanced CT[27,28] or MRI[29]. However, these techniques involve radiological exposure, and are laborious to perform and costly. Recently, Dong et al[30] published a study using radiomic algorithms based on grayscale ultrasound images to elaborate radiomic signatures with the potential to aid in the prediction of VMI, with promising results. Ji et al[31] created predictive models for recurrence after surgical resection using radiomic techniques to analyze contrast-enhanced CT images, with a C-index of 0.633-0.699. In conjunction with the inclusion of clinical data, the model can be used to establish a personalized risk stratification facilitating the individual management of HCC.

Survival after surgical resection has also been assessed in several studies using ML techniques[32-34], and more recently with more advanced DL models based on digitalized histological images of the tumor. Saillard et al[35] drew up a predictive model of survival after resection, attaining a C-index for survival prediction of 0.78. A recent prospective study by Schoenberg et al[36] involving 180 patients also led to a predictive model based on the analysis of 26 preoperative routine clinical variables, obtaining a predictive value of 0.78.

Transcatheter arterial chemoembolization (TACE) is the treatment of choice for intermediate stage B HCC, in the Barcelona Clinical Liver Cancer (BCLC) classification[37]. Adequate selection of patients who might benefit from this treatment is vital in order to avoid unnecessary examinations that can sometimes have undesirable secondary effects for the patient and waste costs for the health system. Studies have been developed based on AI techniques to attempt to predict the response to treatment with TACE and aid adequate patient selection. Most of these studies are based on imaging analysis, though some have used genomic signatures. Morshid et al[38] elaborated a fully automated ML algorithm using the combination of quantitative characteristics of CT images plus pretreatment patient clinical data to predict the response to TACE. They achieved a prediction accuracy rate of 74.2% using a combination of the BCLC stage plus quantitative image features vs using just the BCLC stage alone. Peng et al[39] validated a DL model to predict the response to TACE using CT images from a total of 789 patients in three different hospitals. They obtained an accuracy of 84% and an AUC of 0.97 to predict complete response. Liu et al[40] constructed and validated a DL model (DL radiomics-based C-US model) but based on the quantitative analysis of C-US cine recordings. It was highly reproducible and had an AUC of 0.93 (95%CI: 0.80-0.98) to predict the response to TACE.

Other studies have used ML techniques combining MRI with clinical data to predict the response to TACE. Abajian et al[41] studied 36 patients who underwent MRI before TACE. They developed a predictive model of response with an accuracy of 78%, sensitivity of 62.5% and specificity of 82%.

The efficacy of TACE has also been examined by survival analysis of the patients after its application. Mähringer-Kunz et al[42] developed a prediction model of survival after TACE by constructing an ANN, using all the parameters of the main conventional prediction scores (ART[43], ABCR[44] and SNACOR[45]). They predicted a one-year survival with an AUC of 0.77, sensitivity of 78% and specificity of 81%, better results compared to those of the conventional scores mentioned.

Although most studies evaluating the use of AI to assess TACE have used radiomics, some have also evaluated the prediction of response to TACE using genetic analysis. Ziv et al[46] studied genetic mutations using SVM techniques to predict the tumor response after TACE, though it was a retrospective study with a low number of cases.

Radiofrequency ablation (RFA) as a therapy aimed at curing HCC in early stages[37] has also been evaluated. Liang et al[47] drew up a predictive model of HCC recurrence based on SVM. They studied 83 patients with HCC who underwent RFA, obtaining an AUC of 0.69, sensitivity of 67% and specificity of 86%, with which they were able to identify patients at a high risk of recurrence. Table 2 summarizes the studies using AI techniques in the treatment of HCC.