Hepatocellular Carcinoma (Radiology)

Hepatocellular carcinoma (HCC) imaging relies on a tailored selection of ultrasound (US), computed tomography (CT), and magnetic resonance imaging (MRI), each with distinct roles across surveillance, diagnosis, staging, and treatment planning ^[19]D5.

Ultrasound

Ultrasound is the first-line surveillance modality. Current guidelines recommend semiannual US with α-fetoprotein for at-risk patients ^[3]D5. However, its sensitivity for detecting very-early-stage HCC is limited: a network meta-analysis reported a sensitivity of 0.34 (95% CrI 0.21-0.55) for very-early HCC, no better than AFP alone ^[4]B2a. US performance is further degraded in patients with obesity or steatosis, common in metabolic dysfunction-associated steatotic liver disease (MASLD) ^[2]A1c. For fibrosis assessment, vibration-controlled transient elastography (VCTE) provides validated cutoffs: liver stiffness measurements >10 kPa are associated with increased HCC risk ^[2]A1c. In 2025, the FDA accepted FibroScan LSM as a reasonable alternative to liver biopsy in clinical trials for noncirrhotic MASH ^[2]A1c.

Computed Tomography

Multiphasic CT (noncontrast, arterial, portal venous, and delayed phases) is the cornerstone for HCC diagnosis and staging. Its sensitivity for very-early-stage HCC is 0.68 (95% CrI 0.26-0.93), rising to 0.92 (0.49-0.99) for advanced-stage disease ^[4]B2a. CT is widely used for preprocedural planning, including assessment of vascular anatomy and extrahepatic disease. For transarterial radioembolization ( TARE ), cone-beam CT with perfusion/iodine mapping can approximate microsphere distribution and support personalized dosimetry ^[15]D5. CT also enables future liver remnant volume calculation when resection is considered ^[18]D5.

Magnetic Resonance Imaging

MRI offers the highest diagnostic accuracy for HCC. Contrast-enhanced MRI with extracellular agents (ECA-MRI) achieves a sensitivity of 0.70 (0.50-0.84) for very-early-stage HCC and 0.86 (0.73-0.94) for early-stage disease ^[4]B2a. Hepatobiliary-specific contrast agents, such as gadoxetic acid (Gd-EOB-DTPA), provide additional functional information. The relative enhancement ratio (RER) on hepatobiliary phase (HBP) images, with a cutoff of 0.9, identifies Wnt/β-catenin-activated HCCs that are resistant to immunotherapy (pooled HR 5.79 for poor prognosis) ^[10]B2a. HBP is typically acquired at 20 minutes post-injection ^[10]B2a.

Abbreviated MRI (AMRI) protocols have emerged for surveillance. A prospective two-center study demonstrated that annual dynamic AMRI (D-AMRI), including precontrast, arterial, portal venous, and 3-minute delayed T1-weighted images, plus diffusion-weighted imaging (DWI) and heavily T2-weighted imaging, achieved a diagnostic yield of 6.4% for all-stage HCC compared to 2.2% for biannual US, without increasing the false referral rate ^[11]B2b. The median scan time was 13 minutes, making it feasible for surveillance ^[11]B2b. D-AMRI sensitivity for very-early-stage HCC was 86.4% vs 22.7% for US ^[11]B2b.

For fibrosis assessment, magnetic resonance elastography (MRE) provides accurate staging, with AUROCs of 0.83-0.89 for significant fibrosis and cirrhosis ^[2]A1c. MRI-based proton density fat fraction (MRI-PDFF) is the reference standard for hepatic steatosis quantification, with an AUROC of 0.90 for detecting grade ≥2 steatosis ^[2]A1c.

PET/CT

The role of 2-[¹⁸F]fluoro-2-deoxy-D-glucose PET/CT in HCC is limited due to variable tumor avidity. Evidence for VETC prediction using FDG PET/CT is limited ^[6]B2a. PET/CT may be useful for detecting extrahepatic metastases in selected cases.

Protocol Selection

Protocol selection depends on the clinical indication. For surveillance, US remains the standard, but D-AMRI is a viable alternative in high-risk populations where US is inadequate ^[11]B2b. For diagnosis and staging, multiphasic CT or contrast-enhanced MRI with extracellular or hepatobiliary agents is required. For treatment planning, CT is preferred for anatomic mapping, while MRI with hepatobiliary agents provides functional characterization. For response assessment, enhancement-based criteria ( modified RECIST , LI-RADS TRA) are used ^[15]D5.

The specific imaging findings that establish the diagnosis of HCC, arterial phase hyperenhancement, washout, and capsule appearance, are detailed in the next section.

Pearl: For surveillance, annual dynamic abbreviated MRI (13-min scan) detects three times more very-early-stage HCCs than biannual ultrasound without increasing false referrals ^[11]B2b.

Building on the protocol selection outlined above, the diagnostic interpretation of HCC rests on the Liver Imaging Reporting and Data System ( LI-RADS ) classification, which standardizes the reporting of imaging features in patients at risk for HCC. LI-RADS provides a categorical probability of HCC for each liver observation, with category LR-5 (definite HCC) serving as the noninvasive diagnostic gold standard.

LI-RADS Classification and Diagnostic Performance

The LI-RADS algorithm for CT and MRI (version 2018) uses four major features to classify observations:

• Arterial phase hyperenhancement (APHE) - nonrim, homogeneous or heterogeneous enhancement in the arterial phase.
• Washout - visually assessed temporal reduction in enhancement relative to liver parenchyma from the arterial to portal venous or delayed phase.
• Enhancing capsule - smooth, uniform, sharp border around the observation visible in portal venous, delayed, or transitional phase.
• Threshold growth - ≥50% size increase in ≤6 months or ≥100% increase in >6 months.

Observations are assigned to categories LR-1 (benign) through LR-5 (definite HCC) based on the combination of major features. An individual participant data meta-analysis of 46 studies (6765 patients, 7500 observations) established the pooled positive predictive values (PPVs) for each category: LR-3: 58.28% (95% CI: 44.00, 71.29), LR-4: 80.82% (95% CI: 71.04, 87.86), LR-5: 95.81% (95% CI: 91.06, 98.09) ^[34]B2a. The majority of major feature combinations within the same category had similar PPVs, supporting the current categorization. However, five combinations differed significantly from the pooled PPV: for example, LR-3 observations without APHE measuring <20 mm and without additional major features had a PPV of only 14.81% (95% CI: 6.35, 30.85), while LR-5 observations with APHE measuring 10-19 mm plus threshold growth had a lower PPV of 74.40% (95% CI: 51.06, 89.00) ^[34]B2a.

Ancillary features (AFs) - such as mild-moderate T2 hyperintensity, restricted diffusion, hepatobiliary phase hypointensity, or nodule-in-nodule architecture - are used in LI-RADS to adjust the category. However, a separate individual participant data meta-analysis of the same 46 studies found that application of individual AFs did not impact the area under the receiver operating characteristic curve (AUC) for LI-RADS categories 1-5, nor did it change the PPV, sensitivity, or specificity of LR-5 compared with major features alone (P value range.11 to >.99) ^[32]B2a. This suggests that while AFs may aid reader confidence, they do not improve overall diagnostic performance.

Imaging Mimics of HCC

Several focal liver lesions can mimic HCC on imaging, particularly in the at-risk population. The most important differential diagnoses include:

• Intrahepatic cholangiocarcinoma (ICC) - may show rim APHE, progressive centripetal enhancement, and delayed central enhancement. Diffusion-weighted imaging (DWI) can help: a meta-analysis of 21 studies found that the pure diffusion coefficient (D) from intravoxel incoherent motion (IVIM) imaging was significantly lower in HCC (0.89 × 10⁻³ mm²/s; 95% CI: 0.77-1.02) than in ICC (1.04 × 10⁻³ mm²/s; 95% CI: 0.93-1.16; P <.001), while the apparent diffusion coefficient (ADC) did not differ significantly ^[28]B2a. Radiomics models further improve differentiation: a meta-analysis of 12 studies (2541 patients) reported a pooled sensitivity of 0.82 (95% CI: 0.76-0.86) and specificity of 0.90 (95% CI: 0.85-0.93) for radiomics-based differentiation of ICC from HCC, with an AUC of 0.88 ^[33]B2a.
• Hepatocellular adenoma (HCA) - particularly the beta-catenin-mutated subtype (B-HCA), which carries a high risk of malignant transformation (47.72% in one meta-analysis) ^[29]B2a. B-HCAs often show arterial hyperenhancement (87%), T2 iso-mild hyperintensity (60%), and a central scar (47%), features that overlap with HCC and even focal nodular hyperplasia ^[29]B2a. The presence of necrosis (32.65%) and intralesional hemorrhage may raise suspicion for B-HCA, but biopsy is frequently required for definitive diagnosis.
• Metastasis - typically shows rim APHE and progressive washout, but may be indistinguishable from HCC in the cirrhotic liver. Contrast-enhanced ultrasound (CEUS) with Sonazoid has shown high accuracy for detecting liver metastases (pooled sensitivity 0.88, specificity 0.92) ^[35]B2a, though its role in differentiating metastasis from HCC is less established.

Abbreviated MRI for Screening

Abbreviated MRI (AMRI) protocols, which use a limited number of sequences to reduce scan time while maintaining diagnostic accuracy, have emerged as a screening tool for HCC. A meta-analysis of 19 studies (3914 participants) found that AMRI had a pooled sensitivity of 0.85 (95% CI: 0.83-0.87) and specificity of 0.93 (95% CI: 0.91-0.94) for HCC detection ^[38]B2a. The best-performing protocol was T2-weighted imaging plus DWI plus hepatobiliary phase (T2+DWI+HBP), with a sensitivity of 0.88 (95% CI: 0.83-0.92), specificity of 0.93 (95% CI: 0.91-0.95), and AUC of 0.96 (95% CI: 0.94-0.98) ^[38]B2a. These results support AMRI as a viable alternative to complete contrast-enhanced MRI for HCC surveillance, particularly in patients who cannot tolerate full protocols.

Diagnostic Algorithm

The following algorithm summarizes the imaging-based diagnostic pathway for HCC in at-risk patients:

Step 1: Identify a liver observation on multiphase CT or MRI in a patient at high risk for HCC (cirrhosis, chronic HBV/HCV, NAFLD). Step 2: Classify the observation using LI-RADS major features. Step 3: If LR-5, the diagnosis of HCC is established with a PPV >95% - no biopsy is required. Step 4: If LR-4, consider ancillary features; if they favor HCC, the observation may be upgraded to LR-5 in some guidelines, though evidence shows no improvement in diagnostic performance ^[32]B2a. Biopsy or short-interval follow-up (3-6 months) is appropriate. Step 5: If LR-3, short-interval follow-up (3-6 months) or biopsy is recommended, depending on clinical context. Step 6: If the imaging appearance is atypical or if alternative diagnoses are suspected (e.g., rim APHE suggesting ICC, or a young woman with oral contraceptive use suggesting HCA), use additional sequences (DWI, hepatobiliary phase) or radiomics models to refine the differential. Biopsy remains the reference standard for indeterminate cases.

Diagnostic Test Performance

The table below summarizes the diagnostic performance of key imaging tests for HCC detection and differentiation from mimics, as reported in the meta-analyses cited.

Pearl: The LR-5 category has a PPV of 95.81% for HCC, making it a definitive noninvasive diagnosis - do not biopsy LR-5 observations in at-risk patients; instead, proceed directly to treatment planning.

Beyond the core diagnostic features of arterial-phase hyperenhancement and washout, several additional imaging patterns and signs carry specific prognostic and biological significance in hepatocellular carcinoma (HCC).

Vessels Encapsulating Tumor Clusters (VETC) Pattern

The VETC pattern, sinusoid-like vessels that isolate and encapsulate tumor clusters forming a cobweb-like architecture, is a histologic hallmark of aggressive HCC biology that can be predicted noninvasively on imaging ^[52]B2b. Key imaging correlates include:

• Intratumor necrosis (odds ratio [OR] 6.022; 95% CI 1.427-25.413) ^[52]B2b
• Lower pseudo-diffusion fraction (f) on intravoxel incoherent motion (IVIM) (OR 0.791 per unit increase; optimal cutoff ≤15.7%) ^[52]B2b
• Serum AFP >400 ng/mL (OR 2.962; 95% CI 1.038-8.448) ^[52]B2b
• Non-rim diffuse heterogeneous arterial-phase hyperenhancement and tumor-to-liver signal intensity ratio ≥1.135 on arterial phase or ≤0.585 on hepatobiliary phase ^[52]B2b
• Non-smooth tumor margin and internal arteries on contrast-enhanced CT (CECT) ^[42]B2b

A combined model incorporating f value, intratumor necrosis, and AFP achieved an area under the curve (AUC) of 0.854 for VETC prediction ^[52]B2b. AI-predicted VETC positivity is associated with early recurrence (hazard ratio [HR] 2.34, 95% CI 1.93-2.84) ^[6]B2a.

Targetoid Appearance

Targetoid liver lesions on MRI exhibit a characteristic pattern with differential signal intensity and enhancement between central and peripheral components ^[53]D5. Key features include:

• Rim arterial-phase enhancement
• Delayed central enhancement
• T2 target appearance (central hyperintensity with peripheral hypointensity)
• Targetoid appearance on diffusion-weighted imaging
• Targetoid appearance in the transitional phase or hepatobiliary phase with hepatobiliary contrast agents

This pattern is observed across a wide range of benign, malignant, infectious, inflammatory, and vascular conditions, and a structured diagnostic algorithm integrating imaging findings with clinical context enhances diagnostic accuracy ^[53]D5.

Delayed Central Enhancement and Dilated Vasculature (High FAP Expression)

High fibroblast activation protein (FAP) expression in HCC is independently predicted by two CT features ^[49]B2b:

• Delayed central enhancement (OR 15.196; 95% CI 8.996-25.670)
• Dilated vasculature (OR 7.455; 95% CI 4.928-11.277)

A predictive model incorporating these features achieved AUCs of 0.779 (training) and 0.766 (external validation). High-FAP HCCs are more frequently associated with high-grade intratumoral fibrosis (54.7% vs. 22.4%) and VETC pattern (52.1% vs. 22.6%) ^[49]B2b.

Non-Peripheral Washout

On baseline contrast-enhanced CT, non-peripheral washout in the delayed phase is associated with a lower risk of early progression in patients with unresectable HCC treated with hepatic arterial infusion chemotherapy plus targeted therapy and immunotherapy (OR 0.280; p=0.027) ^[43]B2b.

Mosaic Architecture and Incomplete Capsule

An MRI-based clinicoradiologic model for predicting MVI+/TLSs+ HCC (MT-HCC) incorporates absence of intratumoral hemorrhage, incomplete capsule, and mosaic structure, achieving an AUC of 0.831 in the validation cohort ^[40]B2b.

Non-Smooth Tumor Margin and Internal Arteries

On CECT, non-smooth tumor margin and internal arteries are associated with V/M+ status (VETC and/or microvascular invasion) in solitary early-stage HCC. A combined model including these features plus AFP ≥200 ng/mL and lower tumor-to-liver density ratio in the portal venous phase achieved training AUC of 0.784 and external validation AUC of 0.794 ^[42]B2b.

Washout Time ≤45 Seconds and Marked Washout on CEUS

On contrast-enhanced ultrasound, washout time ≤45 seconds and marked washout are independent predictors for differentiating intrahepatic cholangiocarcinoma subtypes from poorly differentiated HCC ^[46]B3b.

Hyperprogressive Disease (HPD) Pattern

Hyperprogressive disease, rapid progression after PD-1 inhibitor therapy, can be predicted by a transformer-based multimodal model (HOPE) integrating arterial- and portal-phase CT with clinical factors, achieving an AUC of 0.801 in internal validation ^[39]B2b.

Pearl: The VETC pattern is the most prognostically significant imaging-derived vascular signature; a combined model using IVIM f value ≤15.7%, intratumor necrosis, and AFP >400 ng/mL identifies VETC-positive HCC with an AUC of 0.854, stratifying patients at high risk for early recurrence (HR 2.34) ^{[52]B2b[6]B2a}.

From the morphological patterns of established disease, attention now turns to the pre-clinical window, where imaging has the greatest opportunity to alter outcomes. Fewer than 1 in 4 patients with cirrhosis receive any form of HCC surveillance, and most HCCs are diagnosed at a stage beyond curative therapy ^[3]D5. Surveillance is the single most modifiable determinant of early-stage diagnosis.

Who should be screened

Professional societies agree that semiannual surveillance with ultrasound plus α-fetoprotein (AFP) is indicated for all adults with cirrhosis of any etiology ^[3]D5[57]A1c. Among patients without cirrhosis, only selected subgroups with chronic hepatitis B virus (HBV) infection have a sufficiently high annual incidence to justify surveillance; the PAGE-B and REAL-B scores can stratify this risk ^[3]D5. Surveillance is not advised for patients with metabolic dysfunction-associated steatotic liver disease (MASLD) or alcohol-associated liver disease without cirrhosis, because the annual HCC incidence falls below the threshold at which screening is cost-effective ^[3]D5.

Current performance of standard surveillance

Ultrasound and AFP have complementary but limited sensitivity. A network meta-analysis of 170 studies (62 643 participants) reported that ultrasound alone detects very-early-stage HCC with a sensitivity of 0.34 (95% CrI 0.21-0.55), similar to AFP alone at 0.39 (95% CrI 0.28-0.47) ^[4]B2a. Sensitivity improves with advancing stage, but the proportion of early-stage diagnoses remains suboptimal ^[4]B2a. A randomized trial comparing ultrasound alone versus ultrasound plus a panel of biomarkers (AFP, AFP-L3, des-gamma-carboxy prothrombin) found no improvement in early-stage HCC detection (HR 0.81, 95% CI 0.47-1.40) ^[58]A1b.

Risk-stratification tools to improve surveillance efficiency

Because the annual incidence of HCC varies widely across cirrhosis etiologies, one-size-fits-all surveillance is inefficient. The FIB-4 index is the primary first-line risk-stratification tool recommended by the Japanese MASLD 2026 guidelines and other international algorithms ^[2]A1c. Patients with FIB-4 > 2.67, or FIB-4 1.3-2.67 (≥ 2.0 if age ≥ 66 years) combined with abnormal noninvasive liver disease assessment (NILDA), should be referred for hepatology evaluation ^[2]A1c. Platelet count provides complementary information: counts < 150 000/μL indicate high risk, and counts 150 000-200 000/μL indicate intermediate risk ^[2]A1c. For patients with HBV without cirrhosis, PAGE-B and REAL-B scores are externally validated ^[3]D5.

Emerging imaging strategies: abbreviated MRI

Complete liver-protocol MRI is the gold standard for HCC diagnosis but is too time-consuming and costly for population-based screening. Dynamic contrast-enhanced abbreviated MRI (DCE aMRI) protocols can be completed in under 15 minutes and have nearly identical sensitivity and specificity to complete MRI for LI-RADS 5 observations ^[66]D5. Non-contrast MRI (NC-MRI) also shows promise: a meta-analysis of 21 studies (7412 participants) reported a pooled sensitivity of 84% and specificity of 94% for HCC detection, with a diagnostic odds ratio of 78 (95% CI 50-122) ^[62]B2a. The ongoing PREMIUM randomized controlled trial is comparing DCE aMRI plus AFP every 6 months versus ultrasound plus AFP in 4700 patients with cirrhosis, powered to detect a 35% relative reduction in HCC-related mortality (NCT05486572) ^[66]D5. If positive, this trial could establish aMRI as the new standard of care.

Primary prevention: vaccine-based strategies

The most effective strategy to reduce HCC morbidity and mortality is to prevent cirrhosis ^[3]D5. HBV vaccination is a cornerstone of primary prevention; universal infant vaccination programs have dramatically reduced HBV-related HCC incidence in endemic regions. For patients with chronic HBV, antiviral therapy with nucleos(t)ide analogues reduces, but does not eliminate, HCC risk, and ongoing surveillance remains mandatory ^[3]D5. Direct-acting antiviral therapy for hepatitis C achieves sustained virologic response rates exceeding 95%, which lowers but does not abolish the risk of HCC in patients with established cirrhosis ^[61]D5.

Patient education points

Clinicians should communicate that surveillance is underused: fewer than 20% of eligible patients receive it ^[61]D5. Key messages include:

• Semiannual imaging is mandatory, annual surveillance is inferior to 6-monthly surveillance ^[61]D5.
• Ultrasound alone misses many early-stage HCCs; the addition of AFP improves detection modestly, but false positives can occur ^{[4]B2a[58]A1b}.
• Newer options (aMRI, NC-MRI) are being studied and may offer higher sensitivity, but they are not yet standard of care ^{[66]D5[62]B2a}.
• Lifestyle modification reduces risk, weight loss of ≥ 5% improves steatosis, and ≥ 10% may lead to fibrosis regression in MASLD ^[2]A1c.
• Vaccination against HBV is recommended for at-risk adults, and antiviral therapy for viral hepatitis should be optimized ^[3]D5.

Pearl: For every 100 patients with cirrhosis undergoing semiannual ultrasound surveillance, fewer than 40 very-early-stage HCCs will be detected, and fewer than 25 will receive any surveillance at all; aMRI+AFP may change this, pending the PREMIUM trial results ^{[4]B2a[3]D5[66]D5}.

Once treatment is initiated, the radiologist's task shifts from detection to assessing therapeutic response using formal criteria that distinguish residual viable tumor from treatment-related changes. Choosing the wrong criteria set silently misclassifies responders as failures and vice versa, directly affecting decisions to continue, switch, or escalate therapy.

Response Criteria: RECIST 1.1, mRECIST , and LI-RADS TRA

RECIST 1.1 remains the primary endpoint in clinical trials but relies solely on unidimensional tumor diameter measurements ^[67]D5. In HCC, where locoregional and systemic therapies induce necrosis or devascularization without immediate size reduction, size-based criteria underestimate response ^[67]D5. Modified RECIST (mRECIST) addresses this by measuring only the enhancing (viable) component of the target lesion on arterial-phase imaging, classifying complete response as disappearance of any intratumoral arterial enhancement ^[67]D5. However, higher response rates by mRECIST have not consistently translated into improved overall survival, limiting its surrogate validity ^[67]D5.

The LI-RADS Treatment Response Algorithm (TRA) was developed specifically for HCC after locoregional therapy. Version 2017 had limitations: moderate sensitivity for residual viable tumor, ambiguity of the Equivocal category, and overcalling of viable disease after radiation-based therapies ^[78]D5. The v2024 update introduced two separate algorithms, one for non-radiation therapies (e.g., transarterial chemoembolization [TACE ], ablation) and one for radiation-based therapies (e.g., stereotactic body radiotherapy [SBRT ], transarterial radioembolization [TARE ]), reflecting fundamentally different post-treatment imaging evolution ^[68]D5[78]D5.

Performance of LI-RADS TRA v2024

A meta-analysis of 14 studies (1706 patients, 2036 lesions) evaluating LI-RADS TRA v2024 reported a pooled sensitivity of 72% (95% CI: 54-84%), specificity of 95% (95% CI: 78-99%), and accuracy of 83% (95% CI: 78-88%) for detecting residual viable tumor ^[60]B2a. For non-radiation therapies, the pooled prevalence of LR-TR Equivocal was 9% (95% CI: 6-11%), representing a substantial reduction from v2017 ^[60]B2a. For radiation-based therapies, the LR-TR Nonprogressing category, indicating stable or regressing lesions without mass-like enhancement, had a pooled prevalence of 43% (95% CI: 27-60%), appropriately directing continued surveillance rather than immediate retreatment ^{[60]B2a[78]D5}.

A retrospective study of 245 patients directly compared LI-RADS Nonradiation TRA v2024 with v2017 using hepatobiliary agent (HBA)-enhanced MRI as the reference standard ^[74]B2b. v2024 with ancillary features (AFs) achieved the highest sensitivity (79.3%) and accuracy (81.6%), significantly outperforming v2024 without AFs (sensitivity 65.8%, accuracy 71.8%) and v2017 (sensitivity 67.4%, accuracy 73.1%) (all P <.001), with preserved specificity ^[74]B2b.

Modality Choice: CT versus MRI

HBA-enhanced MRI demonstrates superior performance over CT for response assessment after non-radiation locoregional therapy. In the same cohort of 194 patients who underwent both CT and HBA MRI, MRI achieved higher sensitivity (78.7% vs 64.0%; P =.001) and accuracy (81.4% vs 70.6%; P =.002) with no significant difference in specificity (90.9% vs 93.2%) ^[74]B2b. This advantage was most pronounced after TACE (accuracy 82.7% vs 66.1%; P <.001), whereas after ablation both modalities performed similarly (both 79.1%) ^[74]B2b. For radiation-based therapies, CT remains acceptable, but MRI with diffusion-weighted imaging and hepatobiliary phase may improve detection of residual disease ^[15]D5[68]D5.

Timing of Response Assessment

After SBRT, late-term MRI evaluation (9-12 months) provides better prognostic stratification than mid-term assessment (3-6 months) ^[75]B2b. At late-term follow-up, LI-RADS Radiation TRA v2024 was strongly associated with overall survival (HR = 15.80, 95% CI: 3.34-75.00) and progression-free survival (HR = 6.31, 95% CI: 2.68-14.90) ^[75]B2b. Ancillary features did not add prognostic value in this setting ^[75]B2b. For non-radiation therapies, the optimal timing of first post-treatment imaging is typically 1-3 months, with subsequent surveillance at 3-month intervals ^[78]D5.

Emerging Tools: Radiomics and Artificial Intelligence

Radiomics-based machine learning models show promising accuracy for predicting immunotherapy response in HCC. A meta-analysis of 11 cohorts from 9 studies reported a pooled sensitivity of 0.83 (95% CI: 0.75-0.88), specificity of 0.79 (95% CI: 0.68-0.88), and area under the curve (AUC) of 0.88 (95% CI: 0.85-0.90) for predicting response to immune checkpoint inhibitors ^[30]B2a. For TACE, AI models achieved AUCs ranging from 0.55 to 0.97, with radiomics-based models outperforming non-radiomics models (AUC 0.79 vs 0.73) ^[71]B2a. Combined clinic-radiologic models consistently outperformed models based on imaging or clinical features alone ^{[71]B2a[73]B2a}.

Gd-EOB-DTPA-enhanced MRI provides a mechanistic imaging biomarker for immunotherapy response. The relative enhancement ratio (RER) on hepatobiliary phase, with a cutoff of ≥0.9, identifies Wnt/β-catenin-activated HCCs that exhibit an immune-cold tumor microenvironment and resistance to immune checkpoint inhibitors ^[10]B2a. A meta-analysis of 5 studies (253 patients) found that RER ≥0.9 was associated with a 6-fold increased risk of mortality (pooled HR = 5.79, 95% CI: 1.56-21.50) after immunotherapy ^[10]B2a. The number needed to screen with MRI to prevent one early progression by selecting atezolizumab plus bevacizumab over ICI monotherapy was 10 (95% CI: 9.64) ^[10]B2a.

Despite these advances, radiomics and AI tools remain largely investigational. Most studies are retrospective, single-center, and lack external validation ^{[70]B2a[80]D5}. Standardized imaging protocols, transparent reporting, and prospective multicenter trials are needed before clinical adoption ^{[70]B2a[80]D5}.

Controversies and Guideline Disagreement

Pearl: Use LI-RADS TRA v2024 with separate non-radiation and radiation algorithms for locoregional therapy response; for systemic therapy, combine RECIST 1.1 (trial standard) with mRECIST (clinically actionable), and consider Gd-EOB-DTPA MRI RER ≥0.9 as a negative predictor of immunotherapy response (HR 5.79; NNT = 10 to select combination therapy) ^{[10]B2a[67]D5[74]B2b}.

Once treatment response is established, the focus shifts to long-term surveillance for recurrent or de novo HCC, a task complicated by post-treatment architectural distortion. The treated liver may contain lipiodol deposits, ablation zones, radiation fibrosis, or surgical clips that obscure or mimic viable tumor, making modality selection and interpretation critical.

Step 1: Selecting the Surveillance Modality and Interval

After curative-intent resection or ablation, the 5-year recurrence rate approaches 50-70% ^[22]B2a. The AGA 2024 Clinical Practice Update recommends semiannual ultrasound with AFP as the preferred surveillance strategy for at-risk patients ^[3]D5, but post-treatment patients represent an even higher-risk cohort that may benefit from more sensitive imaging.

MRI outperforms ultrasound across all stages. In a prospective study of 407 high-risk cirrhosis patients, MRI with liver-specific contrast achieved an HCC detection rate of 86.0% (37/43) versus 27.9% (12/43) for ultrasound (P < 0.001) ^[81]B2b. Of the detected HCCs, 74.4% were very early stage (single nodule < 2 cm), and 67.4% received curative treatments ^[81]B2b. The 3-year survival of patients with HCC detected by MRI was 86.0%, not significantly different from patients without HCC (94.2%; HR 2.26, 95% CI 0.92-5.56) ^[81]B2b.

Annual abbreviated MRI is a practical alternative. Annual dynamic abbreviated MRI (D-AMRI) showed a higher diagnostic yield than biannual ultrasound for all HCC (6.4% vs 2.2%), early-stage HCC (6.2% vs 2.0%), and very early-stage HCC (4.7% vs 1.2%) without increasing the false referral rate ^[11]B2b. Sensitivity of annual D-AMRI was 84.4% (27/32) versus 28.1% (9/32) for biannual ultrasound (P < 0.001) ^[11]B2b. Median scan time was 13 minutes (IQR 12-15 min) ^[11]B2b.

Non-contrast MRI has limitations after locoregional therapy. Pooled sensitivity of non-contrast MRI (NC-MRI) for HCC detection is 84% with specificity 94% ^[84]B3b. However, for post- TACE monitoring, non-contrast abbreviated MRI (NC-AMRI) showed significantly lower sensitivity for viable tumors in treated lesions (76.5% vs 87.7% for full MRI; P = 0.04) and lower specificity (88.7% vs 96.2%) ^[84]B3b. For overall intrahepatic recurrence per patient, NC-AMRI sensitivity was 78.3% vs 92.4% and specificity 79.6% vs 95.1% (P < 0.001) ^[84]B3b.

Abbreviated MRI protocols in transplant candidates. In cirrhotic patients awaiting liver transplantation , dynamic abbreviated MRI (Dyn-AMRI) and hepatobiliary phase abbreviated MRI (HBP-AMRI) showed sensitivity 93-94% and specificity 94-96%, comparable to full MRI ^[83]B3b. Non-enhanced abbreviated MRI had lower performance (sensitivity 90%, specificity 90%) ^[83]B3b.

Step 2: LI-RADS Treatment Response Algorithm for Recurrence Detection

The LI-RADS Treatment Response Algorithm (LR-TR) v2024 provides a standardized framework for categorizing treated lesions ^[60]B2a. Pooled performance across 14 studies (1706 patients, 2036 lesions) showed:

• Sensitivity 72% (95% CI 54-84%)
• Specificity 95% (95% CI 78-99%)
• Accuracy 83% (95% CI 78-88%) ^[60]B2a

For non-radiation therapies (ablation, TACE), the prevalence of LR-TR Equivocal was 9% (95% CI 6-11%), representing a substantial improvement over prior versions ^[60]B2a. For radiation-based therapies ( TARE , SBRT ), the prevalence of LR-TR Nonprogressing was 43% (95% CI 27-60%), reflecting the slower involution of treated lesions; these non-definitive categorizations appropriately direct continued surveillance rather than immediate retreatment ^[60]B2a.

Ancillary features permitted in the nonradiation treatment response algorithm, mild-moderate T2 hyperintensity and restricted diffusion, can be applied to NC-AMRI, though with the performance trade-offs noted above ^[84]B3b.

Step 3: Special Populations

Post-liver transplant. HCC recurrence after orthotopic liver transplant occurs in up to 15% of patients ^[88]C4. In a single-center series of 329 HCC transplants, recurrence rate was 10.9% (36/329) ^[88]C4. The RETREAT score stratifies risk, but one-third of patients who recurred had RETREAT scores of 0 or 1, limiting its negative predictive value ^[88]C4. The authors recommend a broad screening strategy rather than personalized abbreviated surveillance ^[88]C4. Systemic therapy for post-OLT recurrence is challenging: 10 of 25 patients (40%) had to stop or alter regimens due to adverse effects ^[88]C4, underscoring the importance of early detection.

Post-TARE. Splenic volume increase of ≥18% at 3 months independently predicts early disease progression (< 12 months) with sensitivity 0.74 and specificity 0.97 (AUC 0.86, 95% CI 0.76-0.95), outperforming conventional clinical and laboratory parameters ^[90]B3b. Automated splenic volumetry provides a robust, readily accessible imaging biomarker ^[90]B3b.

Step 4: Emerging Tools for Recurrence Prediction

AI and radiomics. Deep learning models for predicting microvascular invasion (MVI) show pooled sensitivity 0.80 (95% CI 0.78-0.83) and specificity 0.82 (95% CI 0.80-0.85) ^[22]B2a. Contrast-enhanced CT (CECT) models perform best: sensitivity 0.84 (95% CI 0.79-0.88), specificity 0.83 (95% CI 0.77-0.88) ^[22]B2a. AI-predicted VETC positivity is associated with early recurrence (HR 2.34, 95% CI 1.93-2.84) ^[6]B2a. A combined model integrating clinical and CECT features for VETC/MVI prediction in early-stage HCC achieved AUC 0.784 in training and 0.794 in external validation; a derived V/M+ score stratified 2-year recurrence-free survival ^[42]B2b.

Delta radiomics. Dynamic radiomic changes on MRI predict pathological complete response after conversion therapy for initially unresectable HCC. The combined radiomics-AFP model achieved AUC 0.920 in the test set and 0.857 in validation for lesion-level pCR ^[89]B3b.

Blood-based tests. The HelioLiver Dx multi-analyte cfDNA test showed sensitivity 47.8% (95% CI 32.9-63.1) vs 28.3% (95% CI 16.0-43.5) for ultrasound for all HCC; for lesions ≤ 2 cm, sensitivity was 28.6% (95% CI 11.3-52.2) vs 0% (95% CI 0.0-16.1) ^[87]B3b. Specificity was 87.6% (95% CI 85.6-89.4) vs 93.9% (95% CI 92.5-95.2) ^[87]B3b. The test met prespecified co-primary endpoints for superior sensitivity and non-inferior specificity ^[87]B3b.

Step 5: What Not to Do

• Do not rely solely on ultrasound for surveillance in post-treatment patients, especially those with non-viral cirrhosis or obesity. Ultrasound sensitivity for all HCC is only 28.3% and 0% for lesions ≤ 2 cm ^[87]B3b.
• Do not substitute non-contrast abbreviated MRI for full MRI after TACE; NC-AMRI misses a significant proportion of viable tumors (sensitivity 78.3% vs 92.4%) ^[84]B3b.
• Do not use genomic biomarkers or multicancer detection panels in routine surveillance; evidence is insufficient to replace guideline-recommended tests ^[3]D5.

Controversies and Guideline Disagreement

Pearl: After curative treatment, annual dynamic abbreviated MRI detects significantly more early-stage recurrences than biannual ultrasound without increasing false referrals, but full-sequence MRI remains superior for post-TACE monitoring (sensitivity 92.4% vs 78.3% for NC-AMRI) ^{[11]B2b[84]B3b}.

While post-treatment surveillance aims to detect recurrent disease, the interpreting radiologist must first distinguish true progression from the spectrum of iatrogenic and treatment-induced findings that commonly mimic it. Several of these findings represent surgical emergencies that must be recognised on sight, and misreading them as tumour progression can lead to inappropriate management.

Post-Ablation Changes: Thermal Ablation ( RFA / MWA )

Thermal ablation produces a predictable zone of coagulative necrosis surrounded by a thin rim of reactive hyperaemia. On contrast-enhanced CT or MRI, this benign periablational enhancement is typically smooth, concentric, and transient, resolving within weeks to months. Irregular, nodular, or progressive enhancement at the margin signals residual tumour rather than treatment effect.

Complications of thermal ablation include bleeding, infection, and subcapsular liver haematoma ^[91]A1a. In a meta-analysis of CEUS-guided MWA, the pooled complication rate did not differ between combined imaging-biomarker guidance and standard guidance (RR 1.13, 95% CI 0.66-1.91; I² = 0%) ^[91]A1a. Minor complications such as postablation syndrome (fever, malaise, transient pain) and biochemical abnormalities were inconsistently reported across studies, precluding pooled analysis ^[91]A1a. No treatment-related deaths were observed in a multicentre comparison of TACE +MWA versus MWA alone for 3-5 cm HCC ^[100]B3b.

Angiography-assisted cone-beam CT (angio-CBCT)-guided RFA achieved 100% technical success with one Grade 3 complication (7.1%) in a small series; intraprocedural refinements included immediate re-ablation in 21.4% and electrode repositioning in 14.3% ^[103]B3b. Lesion conspicuity improved markedly (median ΔHU 290.1 HU) compared with conventional CT (-10.5 HU) ^[103]B3b.

Post-Ablation Changes: Histotripsy

Histotripsy is a non-thermal, non-ionizing mechanical ablation modality that leaves behind an acellular lysate while preserving collagenous structures >2-3 mm, including blood vessels and bile ducts ^[93]B2a. This unique mechanism produces imaging findings distinct from thermal ablation: persistent enhancement of preserved vascular structures can be mistaken for viable tumour if the radiologist is unfamiliar with the technique ^[93]B2a.

Pooled complication rates from a systematic review (n = 553) were: minor complications (Clavien-Dindo Grade < III) 18.2% (95% CI 4.3%-52.2%; I² = 93.1%) and major complications (Grade ≥ III) 7.0% (95% CI 2.0%-21.5%; I² = 76.4%) ^[93]B2a. Major complications included thrombosis requiring intervention, sepsis, and mortality; no study directly reported treatment-related mortality ^[93]B2a. Case reports describe arterial-biliary fistula and suspected disease hyper-progression after histotripsy, though these events have also been reported after other locoregional therapies and even liver biopsy ^[93]B2a. Off-target radiologic effects (e.g., immune-mediated regression of untreated lesions) occurred at a pooled rate of 10.0% (95% CI 5.0%-20.0%; I² = 6.55%) ^[93]B2a.

Technical success was 94.1% (95% CI 90.4%-96.4%), with evidence of small-study effects (Egger's test p = 0.019) ^[93]B2a. Thirty-day all-cause mortality was 1.3% (95% CI 0.5%-3.5%; I² = 0%), and 90-day mortality was 7.4% (95% CI 3.4%-15.1%; I² = 0%) ^[93]B2a.

Post-Surgical Changes

Laparoscopic hepatectomy (LH) for HCC carries risks of bleeding, bile leak, liver failure, and infection. In a propensity-matched analysis of difficult LH (Iwate score ≥ 7), indocyanine green fluorescence imaging (ICG-FI) significantly reduced severe postoperative complications (Clavien-Dindo ≥ IIIa) from 15.4% to 2.6% (P = 0.048; NNT = 8 to prevent one severe complication), conversion to open surgery from 15.4% to 2.6% (P = 0.048), and postoperative hospital stay (median 8 vs 7 days, P = 0.045) ^[96]B3b. Drainage tube placement duration was also shorter with ICG-FI (7.0 vs 7.0 days, P = 0.048) ^[96]B3b. On imaging, post-surgical changes include fluid collections, bilomas, and perihepatic haematomas that must be distinguished from abscess or tumour seeding.

Post-Radiation Therapy Changes

Radiation-based therapies such as stereotactic body radiotherapy ( SBRT ) and transarterial radioembolization ( TARE ) induce cellular senescence, which manifests on MRI as altered diffusion characteristics and delayed regression of enhancement and tumour volumes ^[104]D5. These physiological changes challenge conventional response criteria ( RECIST / mRECIST ), because persistent enhancement may reflect vascular preservation rather than viable tumour, and volume reduction may be slow ^[104]D5. Functional and temporal imaging biomarkers are needed to differentiate treatment effect from residual disease ^[104]D5.

In a series of MRI-guided RFA combined with ¹²⁵I brachytherapy for HCC adjacent to large vessels (≥3 mm), no seed migration, radiation-induced liver disease, or other complications occurred; technical success was 98.81% ^[102]B2b. Postimplantation dosimetry confirmed adequate coverage (median D90 118 Gy, V100 94%) ^[102]B2b.

Post-Locoregional and Systemic Therapy

Combination regimens of TACE, hepatic arterial infusion chemotherapy (HAIC), immune checkpoint inhibitors (ICIs), and tyrosine kinase inhibitors (TKIs) carry cumulative toxicity. Some patients experience significant adverse effects without therapeutic benefit, underscoring the need for early response prediction ^[98]B3b. A multimodal model integrating CT radiomics, digital subtraction angiography (DSA) features, and clinical parameters achieved AUCs of 0.944 (training), 0.916 (internal validation), and 0.902 (external validation) for predicting early tumour response to TACE-HAIC + ICI + TKI ^[98]B3b.

Surveillance-Related Risks

Surveillance imaging after liver transplantation for HCC adds burdens of cost, incidental findings, and, depending on modality, radiation exposure and contrast-induced complications ^[94]B2b. Prospectively validated binary criteria (outside Milan criteria , AFP ≥ 200 ng/mL, prior hepatectomy with vascular invasion, explant vascular invasion or >3 lesions, or intraoperative extrahepatic extension) identified low-risk patients (76.4% of the prospective cohort) in whom no recurrence occurred after a median 8.0 years of follow-up without surveillance ^[94]B2b. The negative predictive value was 99.2% across pooled cohorts (n = 375), and the number needed to surveil to detect one recurrence was 4.3 in high-risk patients versus 738 in low-risk patients ^[94]B2b.

Pearl: The most common pitfall in post-treatment imaging is misinterpreting benign periablational enhancement, preserved vascular structures after histotripsy, or radiation-induced delayed regression as tumour progression; familiarity with modality-specific temporal evolution and the use of functional sequences (DWI, perfusion) are essential to avoid this error ^{[91]A1a[93]B2a[104]D5}.

Beyond qualitative assessment of treatment-related changes, quantitative imaging signatures extracted from routine contrast-enhanced studies are increasingly used to predict tumor biology and clinical outcomes. These signatures span three tiers: established quantitative biomarkers (e.g., ADC, FDG uptake), radiomic features (handcrafted texture and shape descriptors), and deep-learning-derived representations. The field has matured rapidly, with systematic reviews now available across multiple applications, though clinical translation remains constrained by heterogeneity and limited external validation.

Quantitative Imaging Biomarkers

18F-FDG PET/CT provides the most extensively validated quantitative biomarker in HCC. A meta-analysis of 59 studies (8,585 patients) found that pre-treatment FDG activity was positive in 30.8% of patients and was significantly associated with poorer overall survival (HR 2.05, 95% CI 1.77-2.38), recurrence-free survival (HR 3.18, 95% CI 2.08-4.86), and progression-free survival (HR 1.64, 95% CI 1.36-1.98) ^[111]B2a. The prognostic value held across curative and non-curative treatments and across PET parameters including SUVmax, tumor-to-liver ratio, and visual assessment ^[111]B2a. Although not routinely recommended for HCC staging, FDG PET/CT identifies a biologically more aggressive subgroup and may complement conventional imaging for risk stratification ^[111]B2a.

MRI-derived quantitative parameters such as apparent diffusion coefficient (ADC) and perfusion metrics are not yet standardized as standalone biomarkers. However, radiomic models built on multiphase MRI consistently outperform single-phase approaches. A multicenter study of 249 patients developed a multiphase radiomics model (MP-RM) integrating intra- and peri-tumoral features from arterial, portal venous, and subtraction delta phases, achieving 1-year AUCs of 0.819 (training) and 0.774 (external validation) for predicting early progression after resection or TACE ^[112]B3b. The model effectively stratified high- versus low-risk groups (training HR 3.95, p<0.01; validation HR 1.56, p=0.05) ^[112]B3b.

Radiomics Signatures

Radiomics transforms imaging into high-dimensional feature sets capturing tumor heterogeneity. The most mature applications are summarized in Table 1.

Table 1. Performance of Radiomics and AI Models for Key HCC Endpoints

Microvascular invasion (MVI) is the most intensively studied radiomics target. A meta-analysis of 52 studies (19,531 patients) found that deep learning models achieved pooled sensitivity 0.80 (95% CI 0.78-0.83) and specificity 0.82 (95% CI 0.80-0.85), with a diagnostic odds ratio of 19 ^[22]B2a. Contrast-enhanced CT models performed best among noninvasive modalities (sensitivity 0.84, specificity 0.83, SROC 0.90), while contrast-enhanced ultrasound showed the highest specificity (0.88) but lower sensitivity (0.70) ^[22]B2a. CT-based AI models outperformed radiologists in a separate meta-analysis (AUC 0.89 vs 0.80) ^[24]B2a. However, external validation performance consistently declined: for MVI, external validation yielded sensitivity 0.77 and specificity 0.80 ^[22]B2a; for high-grade HCC, external AUC dropped to 0.75 from 0.85 internally ^[8]B2a.

Vessels encapsulating tumor clusters (VETC) represent a CD34-positive vascular pattern linked to aggressive biology. Radiomics-based AI models achieved pooled sensitivity 0.79, specificity 0.83, and AUC 0.88 across 18 studies (3,615 patients) ^[114]B2a. AI-based methods significantly outperformed non-AI approaches (AUC 0.88 vs 0.82, P=0.01) ^[114]B2a. CEMRI-based AI showed the best performance: sensitivity 0.84, specificity 0.79, AUC 0.87 ^[6]B2a. AI-predicted VETC positivity was associated with early recurrence (HR 2.34, 95% CI 1.93-2.84) ^[6]B2a.

Immunotherapy response prediction is an emerging application. A systematic review of 11 radiomics studies (2,014 patients) found that models predicting response to immune checkpoint inhibitor combinations achieved AUCs of 0.88-0.96 in training cohorts and 0.79-0.88 in validation cohorts ^[70]B2a. Clinical-radiomics models integrating factors such as ALBI grade and portal vein tumor thrombus improved C-indices for overall survival from 0.76-0.77 (radiomics alone) to 0.78-0.82 (combined) in training, and from 0.63-0.69 to 0.67-0.74 in external validation ^[70]B2a. No study used iRECIST criteria, and only one examined ICI monotherapy ^[70]B2a.

Deep Learning Models

Deep learning (DL) extracts hierarchical features directly from images, bypassing handcrafted feature engineering. For HCC detection, a meta-analysis of 41 studies reported pooled sensitivity 87% (95% CI 84-90%), specificity 91% (95% CI 89-93%), and AUC 0.95 (95% CI 0.92-0.96), numerically exceeding human clinicians (AUC 0.91) ^[7]B2a. For MVI prediction, DL models showed significantly higher sensitivity than machine learning models (0.88 vs 0.72, P=0.018) ^[8]B2a.

Multimodal DL models integrating imaging with clinical data consistently outperform unimodal approaches. The HOPE model, a transformer-based architecture combining arterial- and portal-phase CT with clinical factors, predicted hyperprogressive disease after PD-1 inhibitor triple therapy with AUC 0.801 (internal) and 0.687 (external) ^[39]B2b. A multimodal model integrating CT radiomics, hepatic DSA features, and clinical parameters predicted early response to TACE-HAIC plus immunotherapy with AUC 0.902 in external validation ^[98]B3b.

Foundation models and large language models (LLMs) represent the newest frontier. HepatoSageCT, a deep learning foundation model applied to arterial-phase CT, detected esophageal varices with AUROC 0.84 when combined with portosystemic shunt assessment, missing only 4.2% of varices needing treatment (vs 8.4% using shunts alone) ^[113]B3b. It also predicted hepatic decompensation with C-index 0.73 and stratified overall survival (p<0.001) ^[113]B3b. The Fully Automated Stratification System (FASS) integrated automated segmentation (Dice 0.77), radiomic features, and ChatGPT-4o-derived semantic features (irregular margin) to achieve C-index 0.78 for overall survival after resection, with transcriptomic validation showing inflammatory pathway activation in high-risk patients ^[118]B3b.

Limitations and Path to Clinical Translation

Despite promising performance, several barriers prevent routine clinical adoption. External validation is reported in only 16.7% of AI recurrence prediction studies ^[5]B2a and 4 of 11 immunotherapy radiomics studies ^[70]B2a. Performance drops consistently from internal to external cohorts: for MVI, AUC declines from 0.90 to 0.85 ^[22]B2a; for high-grade HCC, from 0.85 to 0.75 ^[8]B2a. Heterogeneity is substantial across studies, with I² values exceeding 70% for most pooled estimates ^{[7]B2a[22]B2a[114]B2a}. Sources include variations in imaging protocols, segmentation methods (73% manual), feature extraction pipelines, and patient populations ^{[22]B2a[70]B2a}. Geographic concentration is a concern: 91.7% of AI recurrence studies ^[5]B2a, 90.9% of immunotherapy radiomics studies ^[70]B2a, and 14 of 15 VETC studies ^[6]B2a originated from China, limiting generalizability to Western populations.

Methodological quality remains suboptimal. The median Radiomics Quality Score across immunotherapy studies was 15/36 (41.7%) ^[70]B2a. Common deficiencies include lack of phantom studies, multi-timepoint imaging, prospective registration, cost-effectiveness analysis, and open data sharing ^[70]B2a. Only one study examined biological correlation of radiomic features ^[70]B2a.

Standardized reporting is urgently needed. For histotripsy, radiologic response assessment varied widely across studies, with some relying on enhancement patterns that may be misleading due to preserved vascular structures ^[93]B2a. The field requires consensus on objective response criteria analogous to mRECIST for thermal ablation.

Pearl: When evaluating a radiomics or AI model for HCC, check for external validation in a geographically distinct cohort, without it, reported AUCs may overestimate real-world performance by 0.10-0.15, as consistently seen across MVI, grade, and immunotherapy prediction studies ^{[8]B2a[22]B2a[70]B2a}.