Hepatocellular carcinoma (HCC) sits at the intersection of liver cancer biology and chronic liver disease. GLOBOCAN 2022 estimated approximately 866,000 incident liver cancers and 759,000 deaths worldwide, which keeps earlier detection and dependable follow-up at the center of HCC care^[1]. Most HCC develops in cirrhosis or other chronic liver disease, so patients often enter an imaging pathway before a tumor is confirmed^[2,3]. In selected at-risk patients, dynamic contrast-enhanced imaging can establish the diagnosis without biopsy, map tumor burden, guide locoregional therapy, and monitor response or recurrence^[4-7].

The problem is not simply whether ultrasound, computed tomography (CT), magnetic resonance imaging (MRI), or positron emission tomography (PET) can produce sharper images. HCC care depends on whether images are acquired at the right phase, reconstructed consistently, quantified reproducibly, and reported in language that clinicians can use. The Liver Imaging Reporting and Data System (LI-RADS), modified Response Evaluation Criteria in Solid Tumors (mRECIST), and treatment response algorithms have helped standardize parts of this pathway^[8,9]. Previous guidelines and modality-specific frameworks have often summarized HCC imaging by individual modality or reporting system^[2-4,8]. That organization is useful, but it can hide the practical question this review addresses: which platform advances actually reduce uncertainty during surveillance, diagnosis, treatment planning, response assessment, or recurrence monitoring?

HCC remains uniquely imaging-dependent because each major clinical decision is conditioned on serial imaging evidence. Surveillance in high-risk populations still begins with ultrasound, often with alpha-fetoprotein (AFP) as an adjunct, despite the well-recognized sensitivity ceiling of ultrasound alone for early-stage disease^[2,3,10,11]. Once a nodule is detected, multiphasic CT, MRI, or CEUS can characterize arterial phase hyperenhancement, washout, capsule appearance, and ancillary features that drive noninvasive diagnosis in at-risk patients^[4,5,8]. The same patient may then undergo MRI for staging, ultrasound or cone-beam CT (CBCT) guidance for therapy, and structured imaging-based response assessment after embolization, ablation, or radiation-related treatment^[6,7,9].

This longitudinal role makes HCC an instructive setting for viewing imaging as a continuum of linked steps rather than as a series of disconnected tests [Figure 1]. Upstream hardware governs whether raw data are sensitive, fast, and stable enough. Acquisition and reconstruction determine whether relevant vascular and tissue contrast is captured, while quantitative extraction and reporting determine whether findings remain comparable across readers and centers. Only then can imaging serve as a reliable basis for treatment planning or response assessment. In practice, the technologies that matter most are those that reduce uncertainty at one or more of these interfaces.

Platform-level advances have not affected all modalities equally, and their value is not defined by novelty alone. Representative strengths and unresolved issues are summarized in Table 1.

Across modalities, the central technical question in HCC is whether acquisition and reconstruction capture the right biology at the right moment. Raw signal alone is not enough. This is most obvious in arterial-phase imaging. Hyperenhancement remains a key diagnostic and response feature, yet it is easy to miss because it depends on bolus timing, temporal resolution, and motion control. Ultrasound partly escapes this problem through continuous real-time sampling, which helps explain why CEUS is so effective for problem solving indeterminate nodules^[16,17]. CT addresses it through speed and reproducible multiphasic timing, increasingly complemented by spectral reconstructions that recover contrast when standard images are borderline^[36-39]. MRI relies less on raw speed than on acceleration, parallel imaging, and motion-robust dynamic strategies that trade sequence complexity for more reliable arterial-phase capture^[51,52,54].

Reconstruction now deserves equal attention because it changes what counts as an interpretable image. Deep learning reconstruction on CT and AI-assisted acceleration on MRI can reduce noise, shorten acquisitions, and rescue studies that would previously have been nondiagnostic^[40,42,54]. The clinical gain is greatest when decision-level performance is preserved, such as maintaining confidence in lower-dose multiphasic CT or making AMRI feasible in frail patients^[41,66]. The complication is that reconstruction can also alter texture, sharpness, or apparent enhancement patterns. Biomarkers that depend on intensity distribution, edge definition, or voxelwise heterogeneity therefore require revalidation when acquisition or reconstruction changes.

These issues explain why acquisition and reconstruction should be treated as part of the clinical evidence rather than as invisible vendor engineering. A new detector, acceleration scheme, or reconstruction network becomes clinically meaningful only when it improves the stability of diagnosis, response assessment, or intervention planning. In HCC, where serial comparison is essential, consistency may matter more than maximal image sharpness.

Quantitative imaging is attractive in HCC because it promises to convert complex vascular and microenvironmental biology into serial, auditable measurements. CEUS time-intensity-curve parameters, CT iodine density, diffusion metrics, stiffness maps, and volumetric enhancement assessments all attempt to move interpretation beyond visual pattern recognition. The recurring problem is that quantitative data are only as useful as their technical stability. Without harmonized acquisition, reconstruction, segmentation, and reporting, thresholds drift across centers and biomarkers remain difficult to qualify for clinical decisions. Relevant standardization frameworks and biomarker classes are summarized in Table 2^{[39,58,62,75-77]}.

It is useful to separate three related but distinct concepts in post-treatment imaging: response measurement, response categorization, and decision impact. mRECIST is primarily a measurement framework. It recognizes that viable HCC is better reflected by arterial-phase enhancement than by the total treated mass, which may remain stable or even enlarge because of hemorrhage, necrosis, edema, or treatment cavity evolution. For this reason, mRECIST remains deeply embedded in HCC trials and routine locoregional therapy follow-up. Its strengths are familiarity, simplicity, and feasibility across multicenter studies. Its limitations are geometric and biologic: a one-dimensional measurement may underrepresent irregular, infiltrative, partially necrotic, multifocal, or heterogeneously treated tumors, and mRECIST does not fully capture extrahepatic progression or mixed intrahepatic responses unless paired with broader disease assessment^[9].

qEASL and related three-dimensional enhancement-based approaches were developed to address this limitation by quantifying viable enhancing tumor volume rather than relying on a single enhancing diameter. This logic is particularly relevant after TACE, transarterial radioembolization (TARE), ablation, or combined locoregional therapy, where residual viable tumor may persist as nodular, crescentic, peripheral, or discontinuous enhancement within a complex treatment zone. qEASL is therefore best framed as a promising quantitative adjunct for selected response-assessment workflows, not as a universal replacement for mRECIST. Its main advantage is biological plausibility: it more directly represents the spatial burden of viable enhancing tumor. Its main barrier is implementation: segmentation, subtraction or registration quality, enhancement thresholds, phase timing, software platform, and local calibration all influence the final value^[78].

LI-RADS treatment response assessment (TRA) provides a complementary solution at the reporting level. Unlike mRECIST or qEASL, LI-RADS TRA is not designed primarily to compute a continuous tumor burden; it standardizes how radiologists categorize treated observations as viable, nonviable, equivocal, nonevaluable, or nonprogressing according to treatment mechanism. The 2024 CT/MRI update is useful because it separates nonradiation and radiation-based treatment response patterns, while CEUS Nonradiation TRA extends structured real-time viability assessment to appropriate nonradiation therapies and resection follow-up. In practice, these frameworks are complementary: mRECIST remains the widely used trial and clinical response metric, qEASL provides a quantitative volumetric adjunct for complex residual enhancement, and LI-RADS TRA provides standardized language for multidisciplinary communication and management decisions^[6,7].

Treatment response imaging in HCC is moving from diameter-only measurement toward mechanism-aware and volume-aware assessment, but these tools are not at the same stage of use. mRECIST is familiar and trial-ready; LI-RADS TRA provides a reporting language for post-locoregional therapy communication; qEASL remains a quantitative adjunct that still needs automated reproducible segmentation, harmonized thresholds, multicenter validation, and prospective decision-impact studies before broad implementation^[6,7,9,78].

Taken together, mRECIST, qEASL, and LI-RADS TRA should be viewed as complementary rather than interchangeable tools. mRECIST remains the most familiar clinical and trial metric, qEASL provides a volumetric enhancement-based adjunct for complex residual viable tumor, and LI-RADS TRA provides standardized post-treatment reporting language for multidisciplinary communication^[6,7,9,78].

Radiomics and AI in HCC are better treated as task-specific tools than as a single technology class. For diagnosis, AI may support lesion detection, arterial-phase timing quality control, LI-RADS feature extraction, and risk stratification of indeterminate observations such as LI-RADS category 3 or 4. For tumor biology, models have been explored for microvascular invasion, histologic grade, cytokeratin 19 (CK19) or Ki-67 expression, molecular subtype, immune microenvironment, and recurrence risk. For treatment planning and follow-up, computational imaging may help estimate ablation margins, tumor territory, residual viable volume, new-lesion detection, and longitudinal response^[79-84].

The most mature near-term application is likely workflow support rather than autonomous diagnosis. Automated or semi-automated liver and tumor segmentation can reduce measurement burden, improve volumetric consistency, and enable qEASL-like or volume-based response workflows. Longitudinal deep learning systems that combine segmentation, tumor volume tracking, and new-lesion detection illustrate how AI could eventually complement RECIST/mRECIST-style workflows. However, such systems should be framed as decision-support and measurement-assistance tools until external validation and clinical impact data show that they improve management decisions or patient outcomes^[84].

The barriers to clinical translation are substantial. Many HCC radiomics studies remain retrospective, single-center, and small relative to the dimensionality of the feature space. Labels such as microvascular invasion, treatment response, recurrence, and progression-free survival are clinically meaningful but can be unstable across institutions because they depend on pathology sampling, treatment selection, follow-up interval, and competing risks. Imaging inputs are also heterogeneous: arterial timing, hepatobiliary phase quality, CT reconstruction kernel, MRI field strength, b-values, contrast agents, and scanner vendors can all shift radiomic features or deep learning representations. Segmentation is another major source of variability, particularly for infiltrative tumors, multifocal disease, necrotic regions, and post-treatment cavities^[85].

For these reasons, the key question is no longer whether an HCC imaging AI model achieves a high internal area under the receiver operating characteristic curve (AUC), but whether it remains calibrated, reproducible, and clinically useful outside the development environment. Essential prerequisites include patient-level data splitting, prevention of data leakage, external multicenter validation, prospective or temporal validation, complete reporting of preprocessing and model selection, assessment of calibration and decision-curve net benefit, subgroup testing across etiology and scanner platforms, and evaluation of how the model changes decisions in a real multidisciplinary workflow. Phantom calibration and cross-vendor harmonization are especially important for quantitative biomarkers and radiomic features; explainability and failure-mode analysis are especially important for deep learning systems^[85-88].

Methodological frameworks should be part of this translation pathway. The Image Biomarker Standardization Initiative (IBSI) provides a foundation for radiomic feature standardization. The Transparent Reporting of a multivariable prediction model for Individual Prognosis or Diagnosis-Artificial Intelligence (TRIPOD + AI) statement supports transparent reporting of AI-based prediction models. The Checklist for Artificial Intelligence in Medical Imaging (CLAIM) 2024 is tailored to medical imaging AI reporting, and the Prediction model Risk Of Bias Assessment Tool-Artificial Intelligence (PROBAST + AI) helps evaluate risk of bias and applicability. These frameworks do not prove clinical utility by themselves, but they define the minimum level of transparency needed before HCC imaging AI can be compared across centers or considered for clinical deployment^[85-88].

At present, radiomics and AI are more defensible as radiologist-supervised workflow tools than as autonomous diagnostic systems. They may help with HCC diagnosis assistance, tumor biology prediction, treatment response evaluation, recurrence prediction, and workflow automation, but most models remain at a research or translational stage. Their near-term value is likely to come from consistency, efficiency, and quantitative traceability inside established reporting workflows^[84-91].

The clinical value of imaging innovation is easiest to appreciate during intervention, where technical gains can be mapped directly to procedural success. Fusion ultrasound and electromagnetic tracking make it possible to target lesions that are inconspicuous on gray-scale ultrasound alone, thereby expanding candidacy for percutaneous ablation^[30-32]. CEUS can then be reused immediately after ablation to assess residual perfusion and to direct same-session retreatment when needed^[7].

For transarterial therapy, CBCT integrates preprocedural and intraprocedural information into a common coordinate system. Feeding-artery identification, tumor territory mapping, and confirmation of target coverage can all be improved within the angiography suite itself^[47-50]. This is one reason HCC is an especially useful model for image-guided oncology: the same technological ecosystem that detects disease also helps deliver and verify therapy.

Response assessment closes the loop. mRECIST remains deeply embedded in trials and clinical practice, but structured CT/MRI and CEUS treatment response algorithms provide a more explicit language for classifying viable, equivocal, or nonviable disease after therapy^[6,7,9]. As volumetric quantification becomes more practical, the next step will likely be better alignment between acquisition, segmentation, response criteria, and retreatment decisions, rather than wholesale replacement of existing frameworks.

Seen from the pathway level, the remaining problem is less a lack of imaging tools than a lack of dependable translation across settings. Access is the first constraint. CEUS agents, spectral CT, 3 Tesla MRI, CBCT-equipped suites, PET/MRI, and new radiotracers are not distributed evenly across regions or reimbursement systems. A technology that works in a tertiary center will not improve HCC outcomes at population scale unless it can be placed into a realistic referral, reporting, and treatment workflow^[2,3,66].

Technical variability is the second constraint. Iodine maps, diffusion metrics, stiffness measurements, and radiomic features all depend on acquisition, reconstruction, segmentation, and software choices. Deep learning reconstruction and new detector platforms can improve images, but they can also change texture, edge definition, or apparent enhancement. Thresholds developed on one scanner or protocol therefore require phantom calibration, cross-vendor harmonization, and transparent reporting before they are used as clinical decision rules^[39,60,85].

Evidence maturity is the third constraint. CEUS, gadoxetic acid-enhanced MRI, MRE for selected risk questions, and CBCT-assisted transarterial therapy have enough clinical experience to guide current practice in appropriate settings^[18,49,62]. Spectral CT has selected-use evidence but still needs broader threshold standardization. Photon-counting CT, FAPI-oriented molecular imaging, IVIM/DKI metrics, and many radiomics or deep learning models should remain labeled as translational until multicenter and prospective studies link them to patient management or outcomes^{[43,44,60,61,72,73,85-88]}.

Governance is the fourth constraint. AI prediction models are often reported with incomplete preprocessing, unclear label definitions, or only internal test performance. The TRIPOD + AI statement, CLAIM 2024, and PROBAST + AI now make these weaknesses easier to identify^[86-88]. For HCC imaging, a credible translation pathway has to evaluate technical performance, clinical utility, workflow fit, and failure modes together.

The next useful advances will probably be those that connect one step of the pathway to the next. Ultrasound can remain the surveillance gateway if visualization scoring, CEUS problem solving, and fusion guidance are used to triage patients rather than to duplicate CT or MRI. CT progress should focus on calibrated spectral and photon-counting measurements for lesions and treated tumors, not only image appearance. MRI development should make hepatobiliary contrast, motion-robust dynamics, diffusion, elastography, and abbreviated protocols easier to deploy in patients who need repeated examinations. PET and hybrid imaging should be reserved for biologic or whole-body questions that CT and MRI cannot answer.

Across modalities, four priorities follow from this logic: multicenter protocol harmonization, prospective decision-impact studies, response frameworks that connect segmentation to retreatment decisions, and AI systems that improve consistency within existing standards. The aim is not to replace expert interpretation with one model or one modality. It is to make HCC imaging more comparable across centers, more explicit about uncertainty, and more useful at the point where management decisions are made^[10].

This review addresses a specific problem in HCC imaging: how to judge whether a platform advance has moved far enough along the pathway to support clinical decisions. The synthesis separates established tools, selected or emerging uses, and translational technologies, and links them to surveillance, diagnosis, treatment planning, response assessment, and recurrence monitoring. That framing makes the maturity gap clearer. CEUS, gadoxetic acid-enhanced MRI, selected MRE, and CBCT-assisted therapy already support clinical care in defined settings, whereas spectral CT, photon-counting CT, FAPI PET, advanced diffusion metrics, radiomics, and many AI models still need standardization and decision-impact evidence. For the wider HCC field, the main contribution is a practical way to align technical development with clinical pathways. Better images matter, but reproducible measurements, common reporting language, and evidence that changes management will matter more for the next stage of HCC imaging.