Table 1.

General characteristics of the included studies.

Ref.	Country	Sample size (total)	Title	Aim of the study	Diagnostic techniques used	AI tool used	Limitations of the study
Liu et al. (2023)	China	30 patients	Diagnosis of primary clear cell carcinoma of the liver based on Faster RCNN	Establish a Faster RCNN model for differential diagnosis of PCCCL and CHCC	Deep learning analysis of CT images	Faster RCNN	Single center study. -The sample size of the patients with PCCCL was small.
Gao et al. (2021)	China	159 patients	Deep learning for differential diagnosis of malignant hepatic tumors based on multiphase CECT and clinical data	Develop an automatic diagnostic model to differentiate types of malignant hepatic tumors	Multiphase CECT	Deep learning model (STIC)	Single-center study. -A limited number of imaging studies. Only typical lesions on MRI were used, excluding lesions with poor quality and more complex lesion types such as infiltrative HCC or complicated cysts. -Pathological proof was not available for all lesions.
Hamm et al. (2019)	USA	296 patients	Deep learning for liver tumor diagnosis part I: development of a CNN classifier for multiphasic MRI	Develop a CNN for classifying hepatic lesions on multiphasic MRI	Multiphasic MRI	Custom CNN	Relatively small training and testing dataset. -The included group was heterogeneous in terms of tumor types. -Insufficient number for some categories, such as focal nodular hyperplasia, liver abscess, liver adenoma, and cholangiocarcinoma. -Using only one type of ultrasound equipment and a single contrast agent for ultrasound. -Valuable information collected in daily clinical practice, such as tumoral markers, was not integrated.
Kim et al. (2020)	South Korea	950 images	Detection of Hepatocellular Carcinoma in Contrast-Enhanced MRI Using Deep Learning Classifier	Develop a deep learning model for detecting HCC using MRI	Contrast-enhanced MRI	Fine-tuned CNN	The image quality of the arterial phase was affected by transient severe motion artifacts. -The training dataset was obtained from a single vendor. -The study population had relatively good liver function. -The model is unable to detect atypical HCCs and low signal intensity in hepatobiliary phase MRI.
Urhuț et al. (2023)	Romania	49 patients	Diagnostic Performance of an AI Model Based on CEUS in Patients with Liver Lesions	Evaluate the accuracy of an automated method for classifying liver lesions using CEUS	CEUS (contrast-enhance ultrasound)	AI system based on algorithms	Single center study. -Lesions segmentation in the training validation set was done manually by doctors.
Nishida et al. (2022)	Japan	55 patients	Artificial intelligence models for the ultrasonographic diagnosis of liver tumors	Construct AI models for diagnosing liver tumors using ultrasonography	B-mode ultrasonography	CNNs based on VGGNet	A single-center retrospective study. -Patients who have specific types of focal liver diseases (small HCC, HCC without pathology, inflammation, etc.) need to be included in future training.
Zhen et al. (2020)	China	201 patients	Deep Learning for Accurate Diagnosis of Liver Tumor Based on MRI and Clinical Data	Develop a DLS for classifying liver tumors based on MRI and clinical data	Enhanced and unenhanced MRI	CNNs based on Inception-ResNet V2	The AI model focuses on diagnosis not detection. -Other types of rare liver tumors were not involved in the training set.