Emerging Role of MRI‐Based Artificial Intelligence in Individualized Treatment Strategies for Hepatocellular Carcinoma: A Narrative Review

ABSTRACT

Hepatocellular carcinoma (HCC) is the most common subtype of primary liver cancer, with significant variability in patient outcomes even within the same stage according to the Barcelona Clinic Liver Cancer staging system. Accurately predicting patient prognosis and potential treatment response prior to therapy initiation is crucial for personalized clinical decision‐making. This review focuses on the application of artificial intelligence (AI) in magnetic resonance imaging for guiding individualized treatment strategies in HCC management. Specifically, we emphasize AI‐based tools for pre‐treatment prediction of therapeutic response and prognosis. AI techniques such as radiomics and deep learning have shown strong potential in extracting high‐dimensional imaging features to characterize tumors and liver parenchyma, predict treatment outcomes, and support prognostic stratification. These advances contribute to more individualized and precise treatment planning. However, challenges remain in model generalizability, interpretability, and clinical integration, highlighting the need for standardized imaging datasets and multi‐omics fusion to fully realize the potential of AI in personalized HCC care.

Evidence level: 5.

Technical efficacy: 4.

Abbreviations

ADC

apparent diffusion coefficient

AI

artificial intelligence

AP

arterial phase

AUC

area under the curve

BCLC

Barcelona Clinic Liver Cancer

CNN

convolutional neural network

DCEI

dynamic contrast‐enhanced imaging

DL

deep learning

DNN

deep convolutional network

DP

delayed phase

DSC

dice similarity coefficients

DWI

diffusion‐weighted imaging

HBP

hepatobiliary phase

HCC

hepatocellular carcinoma

ICI

immune checkpoint inhibitor

LRT

locoregional therapy

LT

liver transplantation

mp‐MR

multiple sequences magnetic resonance

mRCIST

modified response evaluation criteria in solid tumors

MRI

magnetic resonance imaging

mTKI

multi‐tyrosine kinase inhibitor

MVI

microvascular invasion

OS

overall survival

PHLF

post‐hepatectomy liver failure

PVP

portal venous phase

RCIST

response evaluation criteria in solid tumors

RFA

radiofrequency ablation

RFS

recurrence‐free survival

T1WI

T1‐weighted imaging

T2WI

T2‐weighted imaging

TACE

transarterial chemoembolization

TARE

transarterial radioembolization

TLS

tertiary lymphoid structure

VEGF

vascular endothelial growth factor

1. Introduction

Primary liver cancer is the third leading cause of cancer‐related deaths worldwide, with hepatocellular carcinoma (HCC) being the major subtype, accounting for 90% of cases [1]. Over the past decade, the Barcelona Clinic Liver Cancer (BCLC) system has been regarded as one of the most exemplary systems for staging and treatment algorithms for HCC patients [2]. Despite multiple treatment options, HCC remains highly lethal, with a 5‐year survival rate of only 15%. Prognosis varies significantly even among patients at the same stage, likely due to the tumor's complex and heterogeneous pathological and molecular profiles.

Preoperative imaging assessment plays a pivotal role in informing treatment planning by providing noninvasive insights into tumor biology. In HCC, histopathological features such as microvascular invasion (MVI) are closely linked to prognosis, and imaging models reflecting these features have demonstrated strong associations with clinical outcomes [3]. Additionally, liver parenchymal characteristics such as the degree of hepatic fibrosis and the risk of postoperative liver failure are critical in determining treatment eligibility and strategy [4]. Prognostic modeling and response prediction across different treatment modalities further contribute to optimizing clinical decision‐making.

With the advancements in artificial intelligence (AI), the analysis of MRI‐based images is growing rapidly and has become a powerful tool for precise diagnosis and prognosis of HCC. Radiomics enables the high‐throughput extraction of predefined, handcrafted imaging features that characterize tumor phenotype and heterogeneity [5]. In parallel, the rapid evolution of deep learning (DL), which utilizes multilayered neural networks to automatically learn hierarchical feature representations from imaging data, offers a data‐driven and integrative approach to tumor characterization [6]. These AI‐driven techniques are progressively transforming personalized precision management in HCC.

This review outlines the emerging role of MRI‐based radiomics and DL in the noninvasive evaluation of treatment response, prognostic stratification, and therapeutic selection across various treatment modalities. Emphasis is placed on AI‐driven approaches that inform clinical decision‐making regarding treatment response and risk assessment, rather than on methods pertaining to MRI acquisition. To provide a comprehensive perspective, we also summarize the MRI‐based AI application in assessing clinically relevant histopathological features that inform treatment decisions, as well as liver parenchymal status, which lay the foundation for downstream therapeutic decisions. Current limitations are addressed, along with future directions toward the clinical translation of AI‐based MRI tools for individualized HCC management.

2. Literature Search Strategy and Selection Criteria

A structured literature search was performed using PubMed, Embase, and Web of Science to identify studies on the application of AI in MRI‐based individualized treatment of HCC. The search included peer‐reviewed original research articles published in English up to May 2025. Studies evaluating commercial AI software products were excluded. To ensure thematic relevance, the search strategy combined Medical Subject Headings (MeSH) and free‐text terms. Key terms included “Carcinoma, Hepatocellular” OR “HCC”; “Magnetic Resonance Imaging” OR “MRI”; “Artificial Intelligence,” “Deep Learning,” “Radiomics”; and “Treatment,” “Prognosis,” “Recurrence,” “Response,” “Survival,” or “Risk stratification.” Boolean operators (AND, OR) were used to construct the search queries.

Studies were included if they: (1) applied AI techniques to MRI data in HCC patients, particularly in individualized treatment contexts such as prognosis prediction or recurrence risk assessment; (2) were original research articles (retrospective or prospective); (3) were published in peer‐reviewed English‐language journals; and (4) reported sufficient methodological detail, including data sources, model development, and performance metrics (e.g., AUC, accuracy, sensitivity, specificity). Exclusion criteria were:(1) studies not specific to MRI‐based AI applications for HCC; (2) reviews, case reports, editorials, commentaries, or conference abstracts; and (3) articles lacking methodological transparency (e.g., unclear dataset descriptions, lack of specific details for model development).

After database searching, a total of 582 records were identified. Following the automatic removal of 239 duplicates using EndNote 21, two reviewers independently screened the titles and abstracts of the remaining studies according to predefined inclusion and exclusion criteria, resulting in the exclusion of 454 records. Subsequently, 128 potentially eligible articles underwent full‐text review, with any discrepancies resolved through discussion with a third reviewer. Additionally, citation tracking of the included articles was conducted to identify further relevant studies. Ultimately, 114 studies were included in the final analysis (Figure 1).

FIGURE 1.

Flow diagram of study selection.

3. Generally Used AI Models in Clinical Applications

3.1. Radiomics

Radiomics is a computational approach that extracts high‐dimensional, quantitative features from standard medical images, enabling the transformation of imaging data into mineable information [5]. The typical workflow includes image acquisition, tumor segmentation, feature extraction, feature selection, and predictive modeling (Figure 2). By leveraging high‐throughput pipelines, radiomics facilitates the analysis of imaging phenotypes in relation to clinical outcomes, treatment response, and molecular characteristics [7].

FIGURE 2.

Schematic diagram of radiomics analysis.

Currently, radiomics is increasingly applied in various clinical domains, particularly in oncology imaging. It has been used to assist with tumor phenotypic characterization, diagnosis, treatment response assessment, and prognostic prediction [8]. In HCC, radiomics offers a noninvasive method to capture tumor heterogeneity and biological behavior beyond what is visible through conventional imaging. Integrating radiomic features into clinical models has demonstrated improved diagnostic and prognostic accuracy in multiple treatment contexts, supporting more precise and individualized decision‐making in HCC management.

3.2. Deep Learning

DL is an advanced branch of machine learning that utilizes multi‐layered artificial neural networks to automatically learn hierarchical feature representations directly from raw imaging data [6]. This end‐to‐end modeling approach eliminates the need for manual feature engineering and enables direct optimization for clinical prediction tasks. Convolutional neural network (CNN) has long been the dominant architecture in medical imaging, delivering remarkable performance across numerous clinical tasks. More recently, vision transformers and CNN‐transformer hybrid models have gained increasing attention in medical imaging due to their ability to capture global contextual information and model long‐range dependencies more effectively than traditional CNNs [9].

When integrated with complementary data modalities, such as clinical variables or genomic profiles, DL‐based models further enhance the personalization of treatment planning, thereby supporting more informed, patient‐specific therapeutic decision‐making (Figure 3). In liver imaging, DL has facilitated accurate tumor segmentation, supported treatment planning, and enabled noninvasive risk stratification. Architectures such as U‐Net, U‐Net++, and attention U‐Net have shown high precision in delineating liver anatomy and lesions, achieving dice similarity coefficients (DSC) exceeding 90%, and have also been effective in prognosis stratification [10, 11].

FIGURE 3.

Analytical flowchart of a deep learning–based multimodal risk assessment model.

Importantly, DL enables fully automated pipelines, from image preprocessing to outcome prediction, reducing manual intervention and enhancing scalability. Both supervised and unsupervised learning frameworks have been applied to extract high‐level imaging biomarkers and predict clinically relevant endpoints. These capabilities have driven growing interest in DL for HCC applications, including subtype classification, treatment response assessment, and survival prediction, supporting a more precise and data‐driven approach to clinical decision‐making.

4. Image‐Based AI in Pretreatment Decision Support

4.1. Biological Characteristics Refine Treatment Decisions

Histopathological features, particularly microvascular invasion (MVI), are critical indicators of tumor biology and prognosis in HCC. Recent studies have demonstrated the utility of AI‐based MRI in noninvasively predicting MVI and informing adjuvant treatment decisions. For instance, in a multicenter retrospective study involving 206 patients from three institutions, Zhang et al. developed a radiomics model based on T2‐weighted imaging (T2WI), portal venous phase (PVP), and hepatobiliary phase (HBP) images to predict MVI, showing that adjuvant therapy improved survival only in model‐predicted high‐risk patients [12]. Similarly, a multitask DL model trained on a multicenter dataset of 725 patients using T1WI, T2WI, diffusion‐weighted imaging (DWI), arterial phase (AP), and PVP images enabled simultaneous prediction of MVI and RFS. Postoperative TACE was associated with improved RFS among model‐identified high‐risk patients [13].

Despite these advances, most studies have focused on isolated pathological features rather than translating imaging biomarkers into individualized treatment strategies. While some have incorporated prognostic analyses [14], efforts specifically aimed at informing personalized clinical decision‐making remain limited.

4.2. Liver Parenchyma Evaluation

In addition to tumor histopathology, the condition of the liver parenchyma—particularly the degree of fibrosis, risk of postoperative liver failure, and regenerative capacity—plays a critical role in guiding HCC treatment strategies, such as the choice between resection and ablation [4]. These factors are essential in determining surgical feasibility and optimizing the extent of resection.

4.2.1. Liver Fibrosis

Liver fibrosis is one of the important factors affecting the long‐term survival of HCC patients [15]. Accurate assessment is essential for guiding treatment decisions, prognostic stratification, and follow‐up. MRI‐based radiomics and DL approaches have shown promise in fibrosis staging [16, 17, 18, 19, 20, 21]. For example, a single‐center study of 436 patients demonstrated that a radiomics model based on HBP images achieved an Obuchowski index of 0.86, outperforming traditional indices such as normalized liver enhancement (0.77), aspartate aminotransferase‐to‐platelet ratio index (0.60), and the fibrosis‐4 index (0.62) [16]. Moreover, in a multicenter study involving 2063 patients from seven centers, a DL framework using nnUNet‐based liver and spleen segmentation combined with transformer‐derived features from non‐contrast MRI achieved AUCs of 0.810–0.918 in internal validation cohorts and 0.808–0.925 in external validation cohorts for fibrosis staging [20]. Similar results were reported using HBP [18] and T2WI‐based [19] DL models.

4.2.2. Liver Failure

Risk of post‐hepatectomy liver failure is a major consideration when evaluating surgical options. AI models leveraging MRI‐derived morphological and functional parameters have shown reliable performance in predicting liver failure. For instance, in a dual‐center study of 1760 BCLC stage 0/A HCC patients, DL‐based nomograms incorporating liver volume and clinical variables achieved AUCs of 0.78 (whole liver) and 0.81 (remnant liver) for post‐hepatectomy liver failure (PHLF) prediction [22]. Likewise, a radiomics model based on HBP images achieved an AUC of 0.894 for predicting PHLF in a study of 106 patients with HBV‐related cirrhosis undergoing major liver resection [23].

4.2.3. Clinical Relevance

Assessment of liver parenchymal status is critical for tailoring treatment strategies and perioperative management. AI‐based MRI provides a unique opportunity to integrate tumor and liver background characteristics simultaneously, enabling more precise and individualized therapeutic planning. Further research is warranted to validate its clinical applicability.

5. Image‐Based AI in Prognostic Stratification Across Treatment Modalities

In addition to evaluating the biological behavior prior to treatment, AI‐based imaging technologies play a crucial role in prognostic stratification following treatment, offering imperative insights for a personalized treatment plan. In the following sections, we summarize the current applications of imaging‐based AI technologies in the prognostic aspect of personalized treatments.

5.1. Surgical Resection

Surgical resection remains the mainstay curative treatment for early‐stage HCC with preserved liver function [2]. However, postoperative recurrence remains a major challenge, with rates approaching 70% within 5 years. MRI‐based AI models have demonstrated improved accuracy in predicting recurrence and survival, supporting more precise postoperative risk stratification. A summary of MRI‐based AI studies on surgical treatment of HCC is shown in Table 1.

TABLE 1.

Summary of selected MRI‐based AI studies on surgical treatment of HCC.

Study and year	Study design	No. of patients	Treatment	Model type	Modeling	Modality	Segmentation	ROI	Validation strategy	Study endpoint	Result
Kim et al. [24], 2019	Retrospective, single‐center	167	SR	Radiomics	RF	AP, PVP, HBP	3D semiautomatic	Tumor and peritumoral region (3 mm, 5 mm)	Internal validation (temporal split, nested CV, and bootstrap)	DFS	C‐index, validation 0.72
Zhang et al. [25], 2019	Retrospective, single‐center	155	SR	Radiomics	LR	T1WI, T2WI, AP, PVP, HBP	3D manual	Tumor region	Internal validation (temporal split and 10‐fold CV)	ER	AUC: training 0.84; validation 0.84
Wang et al. [26], 2020	Retrospective, dual‐center	201	SR	Radiomics	RF	T1WI, T2WI, DWI, DCE	3D semiautomatic	Tumor region	Internal validation (5‐fold CV)	5‐year survival	AUC: training 0.98; validation 0.76
Chong et al. [27], 2021	Retrospective single‐center	323	SR	Radiomics	SVM	T1WI, T2WI, AP, PVP, TP, HBP	3D manual	Tumor and peritumoral region (10 mm)	Internal validation (5‐fold CV)	ER	AUC: training 0.94; validation 0.84
Zhao et al. [28], 2021	Retrospective single‐center	113	SR	Radiomics	LR	T1WI, T2WI, DWI, DCE	3D manual	Tumor region	Internal validation (randomly split and 5‐fold CV)	ER	AUC: training 0.88; validation 0.87
Zhang et al. [29], 2021	Retrospective single‐center	153	SR	Radiomics	COX regression	T1WI, T2WI, AP, PVP, HBP	3D manual	Tumor and peritumoral region (2 mm, 5 mm)	Internal validation (randomly split and 10‐fold CV)	RFS	AUC: training 0.72; validation 0.73
Wang et al. [30], 2022	Retrospective single‐center	170	SR	Radiomics	RF	ADC	3D manual	Tumor region	Internal validation (randomly split and 5‐fold CV)	ER	AUC: training 0.88; validation 0.85
Cao et al. [31], 2023	Retrospective multi‐center	249	SR	Radiomics	COX regression	DCE	2D manual	Tumor, peritumoral region (10 mm), and liver background	External validation	RFS	C‐index: training 0.89; test 0.85; external 0.80
Feng et al. [32], 2025	Retrospective single‐center	126	SR	Radiomics	Cox regression	AP, PVP, WSI	3D manual	Tumor region	Internal validation (randomly split and 10‐fold CV)	OS	C‐index: training 0.84; validation 0.88
Zeng et al. [33], 2025	Retrospective dual‐center	239	SR	Radiomics	LR	T2WI, DCE	3D manual	Tumor region	Internal validation (randomly split)	ER	AUC: training 0.80; validation 0.74
Guo et al. [34], 2025	Retrospective dual‐center	149	SR and TACE	Radiomics	Cox regression	T2WI, DWI, AP	3D manual	Tumor region	External validation	RFS	C‐index: training 0.82; internal validation 0.82; external validation 0.85
Gao et al. [35], 2022	Retrospective single‐center	472	SR	Radiomics and DL	LR	T2WI, DWI, A DC, DCE	3D manual	Tumor region	Internal validation (randomly split and 10‐fold CV)	ER	AUC: training 0.91; validation 0.84
Yan et al. [36], 2023	Retrospective dual‐center	285	SR	DL	VGGNet19 and GP	AP, PVP, HBP	2D manual	Tumor region	Internal validation (randomly split)	ER	AUC: training 0.95; validation 0.91
Mu et al. [37], 2024	Retrospective single‐center	331	SR	DL	ResNet18	T1WI, T2WI, DWI, ADC, DCE	3D manual	Tumor region	Internal validation (randomly split)	ER	AUC: training 0.84; validation 0.83
Wang et al. [38], 2024	Retrospective single‐center	165	SR	DL	ResNet10	AP, PVP, DP	3D automatic	Tumor region	Internal validation (randomly split)	ER	AUC: test 0.87
Wang et al. [39], 2024	Retrospective single‐center	511	TACE and SR	DL	ResNet18	T1WI, T2WI, DWI	2D automatic	Tumor region	Internal validation (randomly split)	ER	AUC: training 0.87; validation 0.86
Iseke et al. [40], 2023	Retrospective single‐center	120	SR, LT, or TA	DL	VGG16	DCE	3D manual	Whole liver	Internal validation (randomly split and nested CV)	Recurrence	AUC: validation 0.86
Tabari et al. [41], 2023	Retrospective single‐center	97	LT	Radiomics	RF	AP	3D manual	Tumor region	Internal validation (randomly split)	Treatment response	AUC: validation 0.83

Abbreviations: ADC, apparent diffusion coefficient; AP, arterial phase; CV, cross‐validation; DCE, dynamic contrast‐enhanced; DFS, disease‐free survival; DL, deep learning; DWI, diffusion‐weighted imaging; GP, Gaussian process; HBP, hepatobiliary phase; LR, logistic regression; LT, liver transplantation; OS, overall survival; PVP, portal venous phase; RF, random forest; RFS, recurrence‐free survival; ROI, region of interest; SR, surgical resection; SVM, support vector machine; T1WI, T1‐weighted imaging; T2WI, T2‐weighted imaging; TA, thermal ablation; WSI, whole slide imaging.

5.1.1. Radiomics

Owing to the intrinsic advantages of multi‐sequence MRI, radiomic features derived from individual sequences and their combinations have been widely investigated for their prognostic and therapeutic relevance [25, 26, 28, 30, 33, 42, 43, 44, 45]. For example, in a dual‐center retrospective study of 201 patients, a prognostic model using T1WI, T2WI, DWI, and dynamic contrast‐enhanced imaging (DCEI) achieved an AUC of 0.70 for predicting 5‐year survival, with DCEI contributing most to performance [26]. Similarly, in a single‐center study of 113 patients, a multi‐sequence model (including in/out‐phase T1WI, T2WI, AP, PVP, and DP) outperformed single‐sequence models for early recurrence prediction, with the PVP‐based model performing best among individual sequences [28]. Overall, integrating multiple MRI sequences—particularly contrast‐enhanced sequences—consistently improves predictive accuracy.

Region‐of‐interest delineation is a key consideration in radiomics model design. While early studies primarily focused on intratumoral features, growing evidence highlights the value of incorporating peritumoral tissue [24, 27, 29, 31]. For example, in a cohort of 167 patients, adding a 3 mm peritumoral margin yielded comparable accuracy to a postoperative clinicopathologic model for early recurrence prediction (p = 0.76) [24]. Similarly, in 323 patients without macrovascular invasion, a model using tumor and 10 mm peritumoral features achieved an AUC of 0.85 for early recurrence prediction in the validation cohort [27]. Moreover, a multicenter study of 249 patients further demonstrated that combining intratumoral features with either a 10 mm peritumoral zone or background liver improved RFS prediction (C‐indices: 0.77 vs. 0.74 vs. 0.71) [31]. However, the optimal peritumoral range remains undetermined.

In addition, a single‐center study of 126 patients developed a model integrating MRI and pathological radiomics, achieving a C‐index of 0.88 for OS prediction [32]. Despite promising results, the study was limited by small sample size, single‐center design, and lack of mechanistic insight into radiology–pathology interactions.

5.1.2. Deep Learning

Multiple DL models have been developed for prognosis risk stratification in HCC [35, 36, 37, 38, 46, 47], most commonly using CNN architectures such as VGGNet and ResNet [36, 37, 46]. For example, in a single‐center study of 285 patients, multiple sequences magnetic resonance (mp‐MR) DL features extracted from AP, PVP, and HBP images using a VGGNet‐19 architecture were combined with tumor number and MVI to build a DL nomogram, achieving an AUC of 0.91 for early recurrence prediction [36]. Similarly, in a retrospective study of 165 patients, a DL model was constructed using a self‐supervised contrastive and transfer learning from AP, PVP, and DP images [38]. The multi‐model incorporating imaging and clinical features using a modality attention mechanism achieved an AUC of 0.87 and an F1‐score of 87.72% for early recurrence prediction. Another single‐center study of 472 patients constructed a hybrid model integrating radiomics and deep features, achieving an AUC of 0.84 and an accuracy of 0.78 in the validation cohort [35].

5.1.3. Treatment Strategy Optimization

MRI‐based DL models have shown potential in stratifying patients for adjuvant [39] and neoadjuvant therapy [34]. In a single‐center study of 511 patients, a ResNet18‐based DL model achieved an AUC of 0.86 for predicting early recurrence in patients undergoing surgical resection following TACE, suggesting its utility in identifying candidates for conversion therapy [39]. Additionally, DL‐based contrast‐enhanced ultrasound has demonstrated value in selecting between surgical resection and radiofrequency ablation (RFA) [48]. However, MRI‐based AI models for guiding resection versus ablation remain underdeveloped. Future studies may leverage the complementary strengths of MRI and AI to build robust decision‐support tools for individualized treatment planning.

5.2. Liver Transplantation

For patients with early‐stage HCC who are not suitable for surgical resection due to impaired liver function or multifocal tumors, liver transplantation (LT) offers a curative approach by addressing both the malignancy and the underlying liver disease.

5.2.1. Bridging or Downstaging Before LT

Locoregional therapy (LRT), including TACE and ablation, is commonly employed as bridging or downstaging strategies to reduce tumor progression and dropout risk. However, tumor progression despite LRT is associated with poorer post‐transplant prognosis [49], suggesting that dynamic response may reflect tumor biology and inform transplant eligibility. MRI‐based AI models have shown promise in evaluating treatment efficacy in this setting. For instance, in a single‐center study of 97 patients receiving RFA or microwave ablation as initial bridging therapy, Tabari et al. developed a model incorporating AP‐based radiomic features with clinical variables (creatinine, albumin, age, and gender), achieving an AUC of 0.83 for predicting pathological response in the test set [41]. Nevertheless, the role of DL‐enhanced MRI in assessing the efficacy of bridging or downstaging therapies for LT candidates remains underexplored. Future studies are needed to validate such tools for real‐time treatment monitoring and selection optimization in the transplant pathway.

5.2.2. Prognosis Stratification

To further refine transplant eligibility and post‐transplant risk prediction, advanced AI models integrating multimodal data have been developed. For instance, in a recent single‐center retrospective study of 109 patients, a CapsNet‐based AI framework combining clinical variables, radiomic features, and quantitative histopathology achieved an AUC of 0.87 and an F1‐score of 0.84 for predicting post‐LT recurrence, outperforming single‐modality models [50]. Similarly, an XGBoost model using VGG16‐extracted deep features was developed in a cohort of 120 early‐stage HCC patients eligible for LT, yielding AUCs of 0.81 and 0.85 for predicting recurrence at 4–6 years, respectively [40]. Patient‐level analysis further highlighted the added prognostic value of MRI in refining transplant selection criteria.

5.3. Transarterial Therapies

5.3.1. TACE

TACE is the mainstay treatment for patients with intermediate‐stage HCC (BCLC B). Despite being widely used, TACE shows variable efficacy—approximately 60% of patients experience suboptimal responses due to underlying tumor heterogeneity and differential sensitivity to therapy [51]. Early identification of likely non‐responders is critical to facilitate timely treatment modifications and improve outcomes. Relevant MRI‐based AI studies in the context of transarterial and local ablative therapies are summarized in Table 2.

TABLE 2.

Summary of selected MRI‐based AI studies on transarterial and local ablative therapies for HCC.

Study and year	Study design	No. of patients	Treatment	Model type	Modeling	Modality	Segmentation	ROI	Validation strategy	Study endpoint	Result
Kuang et al. [52], 2021	Retrospective, multicenter	153	TACE	Radiomics	LR	T2WI, DCE	3D semiautomatic	Tumor region	External validation	Treatment response	AUC: training 0.83; validation 0.81
Kong et al. [53], 2021	Retrospective, single‐center	99	TACE	Radiomics	LASSO	AP	3D manual	Tumor region	Internal validation (randomly split)	Treatment response	AUC: training 0.86; validation 0.88
Zhao et al. [54], 2023	Retrospective, single‐center	138	TACE	Radiomics	LR	DCE	3D manual	Tumor and peritumoral region (3 mm, 5 mm, 10 mm)	Internal validation (randomly split and 5‐fold CV)	Treatment response	AUC: training 0.91; validation 0.92
Chen et al. [55], 2023	Retrospective, dual‐center	144	TACE	Radiomics	DNN	T2WI, DCE	3D manual	Tumor region	Internal validation (randomly split)	Treatment response	AUC: training 0.97; internal validation 0.83; external validation 0.74
Li et al. [56], 2024	Retrospective, single‐center	124	TACE	Radiomics	RF	DCE	3D manual	Tumor region and peritumoral region (5 mm)	External validation	Treatment response	AUC: training 0.88; test 0.81; external validation 0.81
Song et al. [57], 2020	Retrospective, single‐center	184	TACE	Radiomics	Cox regression	AP, PVP	3D semiautomatic	Tumor and peritumoral region (1 mm, 3 mm, 5 mm)	Internal validation (randomly split and 10‐fold CV)	RFS	C‐index: training 0.74; validation 0.80
Zhou et al. [58], 2024	Retrospective, single‐center	116	TACE	Radiomics	LR	T2WI, AP, PVP	3D manual	Tumor	Internal validation (randomly split)	OS	AUC: training 0.89; validation 0.80
Ince et al. [59], 2023	Retrospective, single‐center	82	TARE	Radiomics	SVM, LR, RF, LGB	AP	3D semiautomatic	Tumor region	NA (nested 5‐fold CV)	Treatment response	AUC: training 0.94
Sozutok et al. [60], 2024	Retrospective, single‐center	65	TARE	Radiomics	LR	T2WI, PVP	3D manual	Tumor region	NA (5‐fold CV)	Treatment response	AUC: training 0.87
Stocker et al. [61], 2025	Retrospective, single‐center	154	TARE	Radiomics	LR	PVP, HBP, ADC	3D manual	Tumor region	NA (elastic net regularization)	Treatment response	AUC: training 0.74
Horvat et al. [62], 2021	Retrospective, single‐center	132	RFA or MVA	Radiomics	RSF	T1WI, T2WI, AP, PVP, HBP	3D manual	Tumor and peritumoral region (5 mm, 10 mm)	Internal validation (random split and bootstrap)	RFS	C‐index: raining 0.98; validation 0.71
Lv et al. [63], 2021	Retrospective, single‐center	58	RFA	Radiomics	LASSO	DCE	3D semiautomatic	Tumor region	Internal validation (random split)	AIR	AUC: training 0.94; validation 0.82
Zhang et al. [64], 2022	Retrospective, single‐center	90	RFA	Radiomics	LR	T1WI, T2WI, DCE	2D manual	Tumor region	Internal validation (random split)	ER	AUC: training 0.87; validation 0.81
Chen et al. [65], 2023	Retrospective multicenter	417	RFA or MWA	DL	ResNet18	T1WI, T2WI, DWI, AP, PVP	3D manual	Tumor and peritumoral region (5 mm)	External validation	LTP	AUC: training 0.87; external validation 1 0.87; external validation 2 0.87
Kong et al. [66], 2025	Retrospective multicenter	289	TA	DL	CNN, RNN	T1WI, DCE	2D manual	Tumor region	External validation	ER	AUC: training 0.93; validation 0.74
Huang et al. [67], 2024	Retrospective, single‐center	164	RFA or MWA	Radiomics	LR	T1WI, DEC, HBP, delta‐radiomics	3D manual	Tumor and peritumoral region (10 mm)	Internal validation (temporal and scanner split and 10‐fold CV)	ER	AUC: training 0.89; temporal validation 0.85; scanner validation 0.83

Abbreviations: ADC, apparent diffusion coefficient; AIR, aggressive intrasegmental recurrence; AP, arterial phase; CNN, convolutional neural network; CV, cross‐validation; DL, deep learning; DNN, deep neural network; DWI, diffusion‐weighted imaging; ER, early recurrence; GP, Gaussian process; HBP, hepatobiliary phase; LGB, light gradient boosting machine; LR, logistic regression; LTP, local tumor progression; MWA, microwave ablation; OS, overall survival; PVP, portal venous phase; RF, random forest; RFA, radiofrequency ablation; RFS, recurrence‐free survival; RNN, recurrent neural network; ROI, region of interest; RSF, random survival forest; SVM, support vector machine; T1WI, T1‐weighted imaging; T2WI, T2‐weighted imaging; TA, thermal ablation; TACE, transarterial chemoembolization; TARE, transarterial radioembolization.

5.3.1.1. Treatment Response

Recent studies have highlighted the utility of MRI‐based AI in predicting response to TACE [52, 53, 54, 55, 68, 69, 70, 71, 72]. For instance, in a single‐center study of 99 patients, a model integrating T2WI‐based radiomics and clinical variables (Child–Pugh grade, BCLC stage, AFP) achieved an AUC of 0.884 for tumor response prediction in the validation cohort [53]. Similarly, a dual‐center study of 140 patients found that a T2WI‐derived radiomics model outperformed DCE‐MRI (AUC: 0.75 vs. 0.60–0.75), and model‐predicted response was significantly associated with OS [55]. In another study of 71 patients, a hybrid model combining radiomics with 3D CNN achieved an AUC of 0.95, surpassing DL‐only (AUC: 0.87) and radiomics‐only (AUC: 0.85) models in identifying progressive disease [71]. Peritumoral features have also shown added value: in a study of 138 patients, combining intratumoral and 3 mm peritumoral radiomics yielded the best performance (AUC: 0.91) [54]. Additionally, radiomic features were also found to be associated with the dynamic changes in viable HCCs at 1 and 6 months following TACE [73].

5.3.1.2. Prognosis Stratification

Beyond treatment response assessment, MRI‐based AI has also demonstrated potential in prognostic prediction and risk stratification for HCC patients undergoing TACE [56, 57, 58, 74]. For example, in a single‐center retrospective study of 184 patients treated with TACE as first‐line therapy, a radiomics model based on PVP tumor regions outperformed models incorporating peritumoral extensions (1–5 mm) in predicting RFS [57]. Multivariate analysis identified Rad‐score, sex, AFP, BCLC stage, tumor margin, and peritumoral enhancement as independent RFS predictors. The integrated nomogram yielded an AUC of 0.80 in the validation cohort.

5.3.2. Tare

Transarterial radioembolization (TARE) has been established as an effective locoregional therapy in the management of HCC. It has been employed with curative intent in early‐stage tumors via radiation segmentectomy, as a bridging or downstaging modality for LT candidates, and as an alternative to TACE in selected patients with intermediate or advanced HCC. Yttrium‐90 (⁹⁰Y) is an increasingly utilized radioisotope in TARE, typically embedded in either resin or glass microspheres.

5.3.2.1. Treatment Response

Advanced AI imaging techniques enable refined assessment of tumor heterogeneity and radiation‐induced changes by extracting high‐dimensional imaging features [59, 60, 61, 75, 76]. For example, in a single‐center retrospective study of 22 patients undergoing ⁹⁰Y TARE, Aujay et al. identified four radiomic parameters—long run emphasis, minor axis length, surface area, and gray‐level non‐uniformity from pre‐ and post‐treatment AP images—as independent predictors of early response per mRECIST [75]. Similar predictive value has been reported for features extracted from T2WI [60] and PVP images [61] based on mRECIST criteria.

In addition to conventional radiomics, recent studies have explored DL approaches. For example, Wagstaff et al. applied a deep convolutional neural network (DNN) to predict 3‐month treatment response to ⁹⁰Y TARE per mRECIST, demonstrating superior performance over voxel‐based dosimetry (F1‐score: 0.72 vs. 0.20; accuracy: 0.65 vs. 0.60) [77]. These models offer potential for improved treatment planning and real‐time response monitoring.

5.4. Local Ablative Therapy

For patients with solitary HCC who are not candidates for surgery, curative ablation offers an effective alternative. However, post‐ablation tissue alterations often result in complex imaging patterns that obscure the differentiation between residual tumor and treatment‐induced changes, limiting the reliability of conventional MRI interpretation.

5.4.1. Treatment Response

In a single‐center retrospective study of 34 patients undergoing RFA, pre‐treatment MRI‐based texture features were first demonstrated to predict complete response, with second‐order features from the Gray Level Dependence Matrix (GLDM) and Gray Level Co‐occurrence Matrix (GLCM) achieving the highest performance (AUC: 0.76–0.78) [62]. However, no subsequent studies have applied MRI‐based AI approaches to evaluate RFA response.

5.4.2. Prognosis Stratification

Radiomic features derived from both tumor and peritumoral regions have shown prognostic value following ablation [63, 64, 67, 78]. For example, in a single‐center retrospective study of 164 patients receiving percutaneous thermal ablation, a delta‐radiomics model incorporating tumor and 10 mm peritumoral regions achieved AUCs of 0.80 and 0.77 in temporal and external validation cohorts, respectively, for predicting early recurrence [67]. Delta‐radiomic features, computed across MRI phases (T1WI, AP, PVP, DP, HBP), underscore the potential of dynamic imaging metrics in guiding post‐treatment surveillance.

DL models trained on preoperative imaging have also demonstrated cross‐modal prognostic utility. Wang et al. developed a Swin Transformer‐based DL model using MRI data from 696 surgically treated HCC patients to predict MVI. When applied to a separate cohort of 180 early‐stage HCC patients undergoing RFA, the model effectively stratified recurrence risk, supporting its role as a noninvasive prognostic tool across treatment modalities [79].

Moreover, post‐ablation imaging features offer additional prognostic information [65, 66]. In a multicenter retrospective study of 289 patients, a PrePost DL model combining pre‐ and post‐treatment multiphase MRI outperformed a preoperative‐only model in predicting local tumor progression (AUC: 0.69 vs. 0.65; F1‐score: 0.47 vs. 0.35) [66].

5.5. Systematic Treatment

Systemic therapy is recommended for patients with unresectable HCC who are ineligible for curative or locoregional treatments, including those with advanced disease (BCLC C), selected intermediate‐stage cases (BCLC B), or progression after LRTs [80]. Available systemic options fall into two main categories: targeted therapies (e.g., multi‐tyrosine kinase inhibitors [mTKIs] and anti‐angiogenic antibodies) and immunotherapies based on immune checkpoint inhibitors (ICIs). Despite therapeutic advances, variability in response and treatment‐related toxicity remain major challenges, highlighting the need for accurate prediction of efficacy to guide personalized therapy. Relevant MRI‐based AI studies focusing on treatment response and prognosis prediction are summarized in Table 3.

TABLE 3.

Summary of MRI‐based AI studies in targeted biomarkers and systematic treatment of HCC.

Study and year	Study design	No. of patients	Treatment	Model type	Modeling	Modality	Segmentation	ROI	Validation strategy	Study endpoint	Result
Gong et al. [81], 2023	Retrospective, sing‐center	107	SR	Radiomics	LR	T2WI, AP, PVP	3D manual	Tumor	Internal validation (random split and 10‐fold CV)	PD‐1/PD‐L1	AUC: training 0.95; validation 0.82
Tao et al. [82], 2023	Retrospective, sing‐center	108	SR	Radiomics	LR	T2WI, AP, PVP	3D manual	Tumor	Internal validation (5‐fold CV)	PD‐L2	AUC: training 0.96; validation 0.87
Long et al. [83], 2024	Retrospective, multicenter	701	SR	Radiomics	LR	T2WI, AP, PVP	3D manual	Tumor and peritumoral region (5 mm)	External validation (nested CV and bootstrap)	TLS	AUC: training 0.92; temporal validation 0.91; external validation 0.91
Li et al. [84], 2025	Retrospective, dual‐center	332	SR	Radiomics	SVM	ADC, AP, PVP	3D manual	Tumor region	External validation (5‐fold CV)	TLS	AUC: training 0.86; internal validation 0.82; external validation 0.88
Long et al. [85], 2025	Retrospective multicenter	600	SR	Radiomics and DL	LR	T2WI, AP, PVP	3D manual	Tumor region	External validation	TLS	AUC: training 0.91; internal validation 0.85; external validation 0.85
Yang et al. [86], 2024	Retrospective, single‐center	124	SR	Radiomics	LR	T2WI, AP, PVP	3D manual	Tumor region	No validation (10‐fold CV)	VEGF	AUC: 0.92
Kang et al. [87], 2025	Retrospective, multicenter	276	TACE and immunotherapy	Radiomics	Six ML a	T1WI, T2WI, DWI, DCE	3D manual	Tumor region	External validation	Treatment response	AUC: training 0.96; external validation 1 0.90; external validation 2 0.89
Dai et al. [88], 2023	Retrospective, multicenter	170	Lenvatinib and anti‐PD‐1 antibody	Radiomics	Neural network	AP, DP	3D manual	Tumor region	External validation (5‐fold CV and bootstrap)	Treatment response	AUC: training 0.99; external validation 0.88
Luo et al. [89], 2022	Retrospective, multicenter	61	TACE and lenvatinib	Radiomics	LR	T1WI, T2WI, DWI, ADC, DCE	3D manual	Tumor region	NA (5‐fold CV)	PFS	AUC: 0.71
Lu et al. [90], 2025	Retrospective, dual‐center	115	TACE, lenvatinib, and PD‐1 inhibitor	Radiomics	RF	T1WI, T2WI, DWI	3D manual	Tumor region	Internal validation (random split and 10‐fold CV)	Treatment response	AUC: training 0.95; validation 0.84
Zhu et al. [91], 2025	Retrospective, dual‐center	102	TACE and PD‐(L) 1 inhibitors	Radiomics and DL	Crossformer and transformer fusion	T1WI, DCE, habitat image	3D manual	Tumor region	External validation (10‐fold CV)	Treatment response	AUC: training 0.87; validation 0.76
Xu et al. [92], 2025	Retrospective, single‐center	111	ICIs and antiangiogenic agents	Radiomics	RSF	DCE	3D manual	Tumor region	Internal validation (random split and 5‐fold CV)	PFS	AUC: training 0.85; validation 0.85

Abbreviations: ADC, apparent diffusion coefficient; AP, arterial phase; CV, cross‐validation; DCE, dynamic contrast‐enhanced; DL, deep learning; DWI, diffusion‐weighted imaging; HBP, hepatobiliary phase; ICIs, immune checkpoint inhibitors; PFS, progression‐free survival; PVP, portal venous phase; RF, random forest; ROI, region of interest; RSF, random survival forest; SR, surgical resection; SVM, support vector machine; T1WI, T1‐weighted imaging; T2WI, T2‐weighted imaging; TACE, transarterial chemoembolization; TLS, tertiary lymphoid structures; VEGF, vascular endothelial growth factor.

^a Decision tree, gradient boosting, support vector machine, logistic regression, K‐nearest neighbors, and random forest.

5.5.1. Immunotherapy

Immunotherapy, particularly anti‐PD‐1/PD‐L1 and anti‐CTLA‐4 agents, has revolutionized the treatment landscape for advanced HCC. Despite the significant success of ICIs in several cancer types, the therapeutic response in HCC remains heterogeneous.

5.5.1.1. Targeted Biomarkers

MRI‐based radiomics models have demonstrated potential for noninvasively predicting immunotherapy‐related biomarkers in HCC. For example, in a single‐center retrospective study of 48 patients, texture features extracted from ADC maps and late arterial phase images significantly correlated with PD‐L1, PD‐1, and CTLA‐4 expression levels assessed by immunohistochemistry [93]. Subsequent studies further confirmed the predictive value of radiomic features for PD‐1/PD‐L1 [81] and PD‐L2 [82] expression.

Beyond individual immune checkpoints, AI‐based MRI models have also been explored for characterizing the tumor immune microenvironment, particularly tertiary lymphoid structures (TLS) [83, 84, 85, 94, 95]. For instance, in a multicenter cohort of 706 patients across four institutions, a radiomic model incorporating intratumoral and 5 mm peritumoral features achieved excellent performance in TLS prediction (AUC: 0.91; F1‐scores of 0.84 and 0.81 in temporal and external validation sets, respectively) and effectively stratified prognosis and response to ICI‐based therapy [83]. A ResNet‐50–based model using T2WI, AP, and PVP sequences also achieved robust performance (AUC: 0.85) in both internal and external test sets for TLS prediction and identified high‐risk patients who benefited from combination immunotherapy [85]. In a dual‐center study of 332 patients, a hybrid model integrating radiomics features (AP, PVP, ADC), BCLC stage, hemorrhage, and satellite nodules achieved AUCs of 0.82 and 0.88 for TLS prediction in internal and external cohorts, respectively, with corresponding F1‐scores of 0.62 and 0.83 [84]. Notably, patients classified as high‐risk by the model derived survival benefit from immunotherapy, supporting its role in treatment stratification.

5.5.1.2. Treatment Response

Beyond target prediction, radiomics‐based AI has also shown value in assessing treatment response [87, 96]. For example, in a multicenter retrospective study of 276 patients with unresectable HCC receiving TACE combined with immunotherapy, radiomic features from six MRI sequences (T1WI, T2WI, DWI, AP, PVP, and DP) significantly improved the predictive performance over clinical variables alone. AUCs increased from 0.70 to 0.90 and from 0.68 to 0.89 in two independent external validation cohorts, respectively, for response prediction per mRECIST [87].

5.5.2. Targeted Therapy

Antiangiogenic targeted therapies include mTKIs (e.g., sorafenib, lenvatinib, cabozantinib, regorafenib) and monoclonal antibodies such as ramucirumab and bevacizumab. Sorafenib, with both anti‐angiogenic and anti‐proliferative effects, has long served as the standard control in first‐line trials for advanced HCC. However, its efficacy is often limited by drug resistance and interpatient variability in response.

5.5.2.1. Targeted Biomarkers

With the advancement of MRI‐based AI, radiomics and DL models have shown promise in predicting targeted therapy‐related biomarkers, such as vascular endothelial growth factor (VEGF), and informing treatment decisions. For example, in a retrospective study of 202 patients with single HCC, Fan et al. demonstrated that multi‐phase contrast‐enhance MRI, particularly PVP and HBP, effectively predicts VEGF expression [97]. A combined model incorporating AFP, irregular tumor margin, and fusion radiomics signatures achieved an AUC of 0.84 in the test set. Similarly, in another retrospective study of 124 patients, radiomics models based on fat‐suppressed T2WI, AP, and PVP sequences yielded AUCs of 0.87, 0.78, and 0.91, respectively, for VEGF prediction [86]. Furthermore, a decision tree‐based model (Radio‐Tree) achieved an AUC of 0.82 [88], highlighting its potential utility in personalizing anti‐VEGF treatment.

5.5.2.2. Prognosis Stratification

Few studies have explored the prognostic role of AI‐based MRI in HCC patients receiving targeted therapy. In a retrospective study of 61 patients with unresectable HCC treated with TACE plus lenvatinib, radiomic features extracted from seven MRI sequences (T1WI, AP, PVP, DP, DWI, T2WI, and ADC) enhanced the performance of a clinical model for predicting disease progression, achieving an AUC of 0.71 [89].

5.5.3. Combined Therapy

The combination of targeted therapy and immunotherapy has been approved as a first‐line treatment of advanced HCC in recent years [98]. MRI‐based radiomics has shown increasing potential in predicting response and prognosis in this setting [90, 91, 92, 99]. For example, in a multicenter retrospective study of 170 advanced HCC patients treated with lenvatinib plus a PD‐1 inhibitor, a radiomics model was developed and externally validated for objective response prediction (per RECIST v1.1), achieving an AUC of 0.82 and demonstrating significant associations with OS and PFS [99]. Likewise, in a dual‐center study of 115 patients with portal vein tumor thrombus receiving TACE, lenvatinib, and a PD‐1 inhibitor, a random forest model based on T2WI, ADC, and AP features achieved the highest AUC (0.79) among four classifiers for short‐term response prediction per mRECIST [90].

6. Current Challenges and Future Prospects

Despite the rapid development of AI‐based MRI in HCC, significant challenges remain in achieving robust clinical translation.

6.1. Methodological Limitations

Many current studies are based on retrospective, single‐center designs and lack external validation, making it difficult to compare model performance and assess clinical utility. As a result, these models have yet to be effectively translated into clinical practice, with currently no prospective studies publicly available to confirm their predictive accuracy. Although nomograms are commonly used to translate radiomics models into individualized decision tools [31, 36, 53, 57, 67, 84, 88], considerable variability exists in their construction, selected features, and validation. Future efforts should prioritize harmonizing model design and validation to facilitate clinical adoption.

6.2. Generalizability

The high sensitivity of AI models to image quality is challenged by variability in scanner types, acquisition protocols, and reconstruction parameters, which compromises model performance across datasets. Addressing these issues by adopting uniform standards—such as those proposed by the Image Biomarker Standardisation Initiative—for image acquisition, segmentation, and processing is essential [100]. Moreover, developing AI models pretrained on medical imaging data, rather than natural image datasets, holds promise for improving both performance and generalizability (Tables 1, 2, 3).

6.3. Interpretability

Limited interpretability remains a major obstacle to the clinical adoption of AI‐based MRI models. While techniques such as SHAP and class activation maps provide partial insights into model decision‐making [38, 39, 47, 90], they often lack biological relevance and do not offer mechanistic explanations for predictions. Consequently, these models are frequently perceived as “black boxes,” reducing clinician confidence and hindering real‐world implementation. To date, only a few studies have explored biological interpretability by linking imaging features to genomic, transcriptomic, or immune markers [43, 83, 85]. Future work should prioritize biologically grounded interpretability to improve transparency and support clinical integration.

6.4. Underexplored Clinical Applications

Several clinically relevant applications of AI‐based MRI remain insufficiently explored. These include the selection of candidates for surgical resection or ablation, personalized systemic therapy selection, and the optimization of post‐treatment imaging surveillance strategies based on tumor biology and patient‐specific characteristics. Additionally, DL applications in prognostic prediction for locoregional and systemic therapies are still limited, mainly due to the data‐intensive nature of these models and the lack of large, annotated medical imaging datasets. Moreover, studies that simultaneously incorporate both tumor features and liver parenchymal status are scarce. Integration of imaging with multi‐omics and spatial pathology is similarly minimal, despite its potential to enhance biologically informed and personalized treatment strategies.

6.5. Future Directions

The development of pretrained AI models based on large‐scale medical imaging datasets and the establishment of standardized, multicenter imaging databases will be critical to enhancing model robustness, reproducibility, and clinical applicability. The promotion of standardized data acquisition, processing, and analysis protocols will facilitate model comparability and scalability. Furthermore, integrating multi‐omics data—such as genomic, transcriptomic, and pathological information—into AI models will provide a more comprehensive understanding of tumor biology. Such integration may support the development of end‐to‐end predictive models that leverage insights from spatial transcriptomics, single‐cell analysis, and molecular profiling, ultimately advancing precision imaging biomarkers for individualized therapy in HCC.

7. Conclusion

AI has markedly advanced MRI‐based imaging in HCC by enabling refined tumor characterization and supporting individualized treatment strategies through high‐dimensional feature extraction. These techniques offer deeper insights into tumor biology and the peritumoral microenvironment. Nonetheless, challenges remain regarding clinical translation, including the need for large‐scale, multicenter imaging databases, biologically interpretable models, and effective multimodal data integration. Addressing these barriers will be critical for transitioning AI‐enhanced MRI from a research tool to a clinically actionable component of precision oncology in HCC management.

Che F., Zhu J., Li Q., Jiang H., Wei Y., and Song B., “Emerging Role of MRI‐Based Artificial Intelligence in Individualized Treatment Strategies for Hepatocellular Carcinoma: A Narrative Review,” Journal of Magnetic Resonance Imaging 63, no. 1 (2026): 79–97, 10.1002/jmri.70048.