From August to November 2026, AI research on non-invasive assessment of liver fibrosis—centered on metabolic dysfunction–associated steatotic liver disease (MASLD) and its severe form, metabolic dysfunction–associated steatohepatitis (MASH)—was densely published: among thirteen studies, eleven were recorded within these four months; one was a follow-up study published in March of the same year; and another listed only the year. Two threads intertwine. One revisits an old question: Can predictive models outperform FIB-4—a simple index calculated from routinely available clinical data? None of the three lines of evidence comparing ‘models versus FIB-4’ demonstrates that new models are stronger: the pooled AUC difference from a validation-audit-style systematic review crosses zero (ΔAUROC = 0.086, 95% CI −0.448 to 0.621); the point estimate for CT-based body composition scoring lags behind by 0.03 (p = 0.37); and a decision-support prototype, tested on the same locked test partition, yields 0.81 versus FIB-4’s 0.92. The other thread—less conspicuous but more robust—is that the act of measurement itself is being redone. AI first standardizes ultrasound region-of-interest (ROI) placement (inter-physician consistency: 81%, κ = 0.25), then prospectively identifies viscosity—not stiffness—as the strongest independent determinant of ‘liver stiffness’, and finally transforms histopathological slides into continuous readings beyond categorical staging (as summarized herein).
## Three pieces of evidence—none proves ‘superiority’
A validation-audit-style systematic review and meta-analysis published in *Cureus* (teams from South Tees Hospitals NHS Foundation Trust, King’s College Hospital, and others; six authors; ICMJE declaration states no financial relationships among all authors) asks a narrow question: For advanced fibrosis (≥F3), when predictive models and FIB-4 are head-to-head evaluated on the same cohort of patients and same endpoint, is the model superior? The team searched four literature databases (PubMed, Europe PMC, OpenAlex, Semantic Scholar), yielding 23,420 deduplicated records; 2,892 articles met eligibility criteria for assessment; 44 potential within-cohort comparisons were identified; and only three head-to-head studies ultimately satisfied inclusion criteria after final verification. The primary analysis included two separate-cohort biopsy-validated studies (source category labels retained verbatim; study-level ΔAUROC values were 0.053 and 0.140), yielding a pooled ΔAUROC of 0.086 (95% Hartung-Knapp CI: −0.448 to 0.621). Adding one internal hold-out set study (ΔAUROC = 0.165) raised the pooled value to 0.126 (95% CI: −0.023 to 0.276). All three point estimates are positive, yet both pooled confidence intervals include ‘no difference’. The review also performed a spot audit on 500 automatically excluded records and found no false negatives (negative predictive value = 100.0%, 95% CI: 99.2%–100.0%). Two structural findings matter more than effect sizes (per this article’s judgment): head-to-head evidence is too sparse and too imprecise; and all included evaluations were conducted solely by the original model development teams—no independent third party was involved. The conclusion is quoted verbatim: current evidence does not support replacing FIB-4 with predictive models; future research should prioritize independent, multicenter validation, calibration, clinically actionable thresholds, and direct assessment of net clinical benefit. This article judges: the funnel—from 23,420 records down to just three eligible studies—indicates the bottleneck in this field is not model output volume, but rather the number of ‘same-cohort, same-endpoint’ comparisons; although all point estimates are positive, the wide confidence intervals (e.g., −0.448 to 0.621) reflect ‘insufficient evidence’, not ‘proof of inefficacy’—these two interpretations must not be conflated.
An interpretable CT-Fibrosis Score was published in *Abdominal Radiology* (University of Pittsburgh team; ten authors; declared no competing interests): 151 biopsy-confirmed MASLD patients (86 with advanced fibrosis F3–F4, 65 with early-stage F0–F2) were retrospectively enrolled; deep learning segmented and quantified abdominal fat, skeletal muscle, bone, liver, and spleen metrics; stratified five-fold cross-validation was repeated 50 times, with imputation, standardization, and feature selection strictly confined to training folds; hyperparameters were selected via inner five-fold cross-validation; sensitivity analyses covered missing data, age dependence, and ascites. Four consistently selected features: age, bone-to-abdominal volume ratio, visceral adipose tissue density, and splenic index. Out-of-fold AUC was 0.77 (95% CI: 0.69–0.85). Crucially, two comparisons were made (on 141 cases with available FIB-4): CT score 0.76 vs. FIB-4 0.79, difference −0.03 (95% CI: −0.11 to +0.04; p = 0.37); authors explicitly note this comparison was not powered for non-inferiority. When the CT score was added to FIB-4, AUC rose to 0.80 vs. FIB-4’s 0.79, difference +0.003 (95% CI: −0.04 to +0.05; p = 0.90; 0.80 and 0.79 are source-reported rounded values; +0.003 is the source-reported difference—this article reproduces it without recalculation); the confidence interval excludes any improvement greater than ~0.05. Authors’ conclusion, quoted verbatim: discrimination is moderate; adding CT score to FIB-4 yields no incremental gain; external validation in unselected populations undergoing CT for non-hepatic indications is required before any ‘opportunistic application’. This article judges: this is the most complete negative-result template among these publications—within-fold fitting prevents leakage, statistical direct comparison uses DeLong testing, an ‘upper bound on incremental gain’ is reported, and absence of non-inferiority design is explicitly noted; the correct interpretation of a negative result is not ‘proof of equivalence’, but rather ‘the possible improvement upper bound is already very small’.
A hepatology decision-support prototype, HepDecide, was published in *Healthcare* (sole author from Istanbul Health and Technology University; declared no conflicts of interest), integrating stage-specific tabular classifiers, SHAP explanation layers, and retrieval-augmented generation anchored to EASL and AASLD guidelines, outputting HL7 FHIR R4 resources, with input mapped to LOINC and diagnoses to ICD-10. Its self-audit is most noteworthy (per this article’s judgment): because patient-level predictions from original development runs were not retained, the author re-evaluated the prediction layer under strict leakage-control conditions: using a locked, patient-wise test partition; performing imputation, scaling, and resampling only within training folds; testing 25 classifier–imbalance configurations per endpoint; calibrating exclusively on training data; and conducting calibration, precision–recall, decision-curve, and SHAP analyses. Locked test-partition results on two publicly de-identified cohorts (biopsy-characterized MASLD cohort, n = 605; HCC cohort, n = 165): steatohepatitis ROC-AUC = 0.67 (95% CI: 0.50–0.84, test n = 121); cirrhosis (F4) = 0.81 (0.52–1.00, n = 121, only 5 events); 1-year HCC survival = 0.76 (0.56–0.92, n = 33); Brier scores = 0.096, 0.028, 0.199 respectively; development-phase estimates of 0.90, 0.98, and 0.92 are separately archived (source notes original input tables and individual test probabilities for cirrhosis are no longer retrievable; correspondence between archived values and endpoints is inferred by this article from source-listed endpoint order—source did not annotate each individually). Direct head-to-head comparison on the same locked cirrhosis test partition: FIB-4 yielded ROC-AUC = 0.92 (0.82–0.99). Top-five SHAP-ranked features accounted for 58%–73% of total attribution; the language-generation layer, due to lack of preserved logs, receives no quantitative evaluation statement. Authors’ conclusion, quoted verbatim: under leakage control, the prediction layer shows moderate discrimination and wide confidence intervals; it did not outperform FIB-4 for cirrhosis; the prototype’s value lies in standardized, interpretable, guideline-traceable design and fully documented, reproducible evaluation procedures; it is positioned as a ‘triage-and-reporting prototype under clinician supervision’, not a clinically deployable system. This article judges: the gap between 0.98 (development-phase estimate) and 0.81 (leakage-controlled re-evaluation) reflects the distance between ‘development-phase estimation’ and ‘leakage-controlled re-assessment’ for the same endpoint; the cirrhosis endpoint’s confidence interval—0.52 to 1.00—reflects a stark sample-size warning; and the author’s archiving of both sets of numbers confers methodological value exceeding the numbers themselves.
## Two positive narratives—answering different questions
The review’s conclusion does not mean no one has offered positive narratives. Within this same timeframe, two studies each claim superiority—but neither wins the question ‘Can it replace FIB-4?’
A retrospective analysis published in *Hepatology Communications*, based on 1,970 biopsy-confirmed patients screened from the resmetirom phase III trials (MAESTRO-NASH, MAESTRO-NAFLD-1) (teams from UC San Diego MASLD Research Center, Madrigal Pharmaceuticals, et al.; six authors; conflicts of interest quoted verbatim: three authors are Madrigal employees and shareholders; one former consultant; first author lists consulting relationships with >20 pharma companies and institutional research funding from Madrigal; one author declares no conflicts), built a random forest model using 37 clinical, laboratory, and imaging predictors to forecast fibrosis stage and ‘MASH with F2/F3 fibrosis’. Full-model (37-variable) one-vs.-rest four-fold cross-validation average AUCs were 0.76, 0.75, and 0.94 (for F0/F1, F2/F3, F4 respectively); best ‘Lean’ model (subset of routinely available clinical variables) yielded 0.83, 0.78, and 0.89; for MASH with F2/F3, full-model AUC = 0.75, Lean model = 0.79. Liver stiffness measurement (LSM) ranked among top predictors alongside FIB-4 and FibroScan-AST; the Lean model incorporating LSM outperformed ‘blood-biomarker-only’ models and also surpassed the ‘FIB-4 + LSM’ combination; top-k minimalist models retaining LSM preserved most performance. Authors set their own boundary: additional external and prospective validation is required before clinical deployment. Three details warrant separate mention (per this article’s judgment): First, ‘superiority’ is claimed over the ‘FIB-4 + LSM’ combination—not FIB-4 alone—so this does not address the same question posed in the prior review: ‘Can it replace FIB-4?’ Second, the Lean model outperforms the full model for F0/F1 (0.83 vs. 0.76) but underperforms for F4 (0.89 vs. 0.94), indicating variable reduction is not uniformly detrimental. Third, this is a trial-screening population—fully biopsy-confirmed and enriched by inclusion/exclusion criteria—so performance cannot be directly extrapolated to general outpatient settings (this sentence represents this article’s minimal inference based on the source’s ‘trial screening’ description).
A plasma proteomic score, MASHPredict, was published in *Frontiers in Immunology* (teams from Affiliated Huai’an Hospital of Xuzhou Medical University, et al.; 13 authors; declared no commercial or financial interests): 142 participants from a single Chinese hospital (83 MASH, 59 controls) were randomly split 7:3 into training and validation sets; LASSO regression combined with Boruta algorithm selected features; ten machine learning models were compared; two external test cohorts were additionally defined. AUCs were: training = 0.83 (95% CI: 0.74–0.92), validation = 0.80 (0.66–0.94), test cohort 1 = 0.79 (0.67–0.91), test cohort 2 = 0.79 (0.76–0.83); the paper claims superiority over FIB-4, APRI, AST/ALT ratio, and NAFLD Fibrosis Score. Prognostic results quoted verbatim: high-risk group hazard ratios (HRs) for liver-related death = 8.18, MASH = 12.07, cirrhosis = 3.37, HCC = 3.11, ASCVD-related death = 1.64, all-cause death = 1.42 (all p < 0.001); authors position it as a validated non-invasive diagnostic tool supporting early detection, risk stratification, and clinical decision-making. This article judges: the 142-person development sample and single-center origin are obvious limitations; the composition of the ‘two external test cohorts’ remains unelaborated in the abstract; the claim of ‘superiority over FIB-4’ must be read within the structural context provided by the prior review: all head-to-head assessments included in that review were conducted solely by original model developers—such superiority claims lack precisely the independent third-party replication (this article does not question authenticity, only flags the evidentiary structure).
## Imaging staging: the drop from training to external testing
In the same week, *Abdominal Radiology* also published a staging study using topological deep learning (Qingdao University Affiliated Hospital team; eight authors; declared no competing interests): TopoEfficientViT integrates topological data analysis with the EfficientViT framework to stage fibrosis from hepatobiliary-phase MR images. Source motivation quoted verbatim: early fibrosis is reversible, so accurate staging is critical; biopsy remains the gold standard diagnosis, but its invasiveness limits clinical use. The study retrospectively enrolled 415 chronic liver disease patients from June 2018 to July 2025 (323 male, mean age 56.52 ± 9.54 years; 92 female, 58.86 ± 9.68 years); data from Site A (same institution) were split into training (n = 275) and validation (n = 75) sets; Site B served as the test set (n = 65). AUC and accuracy for F1–2 vs. F3–4 staging: training = 0.942 and 0.906, validation = 0.849 and 0.76, test = 0.829 and 0.769; for F1–3 vs. F4: training = 0.957 and 0.891, validation = 0.861 and 0.813, test = 0.791 and 0.785 (source abstract contains typographical error ‘0.861and’—no space; reproduced verbatim). This article judges: the drop from training AUC 0.957 to external test AUC 0.791 equals 0.166 (recalculated herein); moreover, Site B and Site A belong to the same institution, so its ‘external’ nature is weaker than true cross-institutional validation.
## Measurement reproducibility: standardizing ROI placement
A study published in *European Radiology* addresses the measurement side: In liver shear-wave elastography (SWE), where should the circular region of interest (ROI) be placed—and who decides? (Team from Massachusetts General Hospital Ultrasound Research and Translational Center; 18 authors; conflicts of interest quoted verbatim: two authors disclose relationships with GE HealthCare; three co-authors are GE HealthCare employees; the external validation dataset was provided by GE HealthCare; model development and analysis were conducted by non-industry co-authors.) Their two-stage pipeline first performs self-supervised pretraining using SimCLR on 17,441 unlabeled ultrasound images, then fine-tunes an ENet segmentation model on 495 labeled liver images and 1,564 images with poor probe contact; the segmentation output, constrained by ‘effective hepatic tissue pixels, axial depth, lateral boundaries, and regions of poor acoustic coupling’, synthesizes an optimal ROI mask; a distance-transform algorithm locates ROI coordinates; the same pipeline also measures skin-to-liver capsule distance and provides guidance on probe angle and contact. The reader study comprised 18,182 images from three ultrasound systems, 654 liver elastography exams, and 609 patients (mean age 52.6 ± 12.95 years, 55.4% male): liver segmentation Dice = 0.89, sensitivity = 0.95, specificity = 0.93, AUC = 0.94; AI–physician ROI placement agreement = 93.8%; inter-physician agreement = 81%; κ between two attending physicians = 0.25. This article judges: κ = 0.25 is the clearest number across these publications—the long-discussed ‘inter-operator variability’ in SWE is here directly quantified; AI’s role here is not diagnosis, but measurement. Boundaries must also be clarified: this study did not assess how standardized acquisition affects final staging accuracy (abstract does not declare this endpoint); it improves reproducibility—not accuracy.
## Viscosity, not stiffness, is the strongest determinant of ‘stiffness’
A prospective study published in Academic Radiology raises earlier questions about input validity (ultrasound department, Beijing Friendship Hospital, Capital Medical University; six authors; declared no competing interests): 573 adults with MASLD underwent multiparametric ultrasound to measure shear wave elastography (SWE) stiffness, ViPLUS viscosity, and attenuation (ATT); participants were stratified by BMI <24 versus ≥24 kg/m², and inverse probability treatment weighting (IPTW) was applied to balance covariates; after adjustment, prespecified clinical covariates were well balanced between the two groups. Results: Hepatic viscosity was the strongest independent determinant of liver stiffness (β = 5.80, 95% CI 5.33–6.27, p < 0.001), equivalent to an approximate 0.58 kPa increase in SWE stiffness per 0.1 Pa·s rise in viscosity; this dominance held across both BMI strata—and was stronger in the high-BMI group (≈0.63 vs. 0.51 kPa/0.1 Pa·s). XGBoost and SHAP were used solely for exploratory feature attribution (authors explicitly stated they were not developing an independent diagnostic or predictive tool); ViPLUS emerged as the strongest model feature, consistent with regression findings; AST/ALT ratio and HDL cholesterol showed nonlinear associations, whereas triglycerides and glucose showed no significant nonlinearity; some low-BMI individuals also exhibited high-risk hepatic phenotypes. Authors’ conclusion verbatim: 'Liver stiffness reflects viscosity-related viscoelastic changes—not fibrosis alone'; multiparametric ultrasound integrating viscosity and stiffness may offer incremental identification value for high-risk individuals, including those with low BMI. This article’s assessment: This is a validity audit at the input level. Prior statements such as 'LSM ranks among the most important predictors' and ROI standardization both target stiffness-type measurements; if stiffness primarily carries inflammation-associated viscosity signals, then 'elevating AUC by adding LSM to models' may reflect improvement in detecting a mixture of inflammation and fibrosis—not fibrosis alone—and fluctuations in stiffness values before and after therapeutic intervention must be reinterpreted accordingly (this article infers this direction as an extension of the source’s conclusions).
## Turning Pathology Slides into Continuous Readouts
At the other end of measurement lies the pathology slide—the reference standard against which all noninvasive methods are compared. Virchows Archiv published Tri-MASH, a trichrome-stain deep learning model (teams from the University of Minnesota and Mayo Clinic, among others; 20 authors; conflict-of-interest disclosure verbatim: one author was employed by Aiforia Technologies during the study; the study used Aiforia’s commercial platform; Aiforia did not sponsor the research nor control analysis or submission decisions; cohort two slides originated from the Emricasan trial, with publication permission granted by current rights holder NOBO Medicine). Tri-MASH quantitatively assesses steatosis, classic hepatocyte ballooning, and fibrosis on single slides and analyzes their spatial relationships; inflammatory cell density was an exploratory metric. Training used 161 specimens; internal validation used 30 nonoverlapping specimens; external cohort one comprised 39 biopsies independently scored by four pathologists; cohort two comprised 524 baseline and week-72 biopsies from 262 clinical trial participants (262 × 2 = 524, recalculated herein). Internal validation F1 scores (a classification performance metric, unrelated to fibrosis stage F1) ranged from 78.9% (inflammatory cell detection) to 99.2% (perisinusoidal fibrosis); correlations with pathologist median assessments in cohort one: steatosis Rs = 0.85, fibrosis Rs = 0.87–0.88, ballooning Rs = 0.55; in cohort two, exploratory pooled analysis correlations dropped to steatosis 0.52, ballooning 0.54, total fibrosis and portal septal fibrosis 0.50 (source notes confidence intervals and p-values were not corrected for within-participant repeated observations). Inflammatory cell density correlated poorly with conventional lobular inflammation grading (Rs = −0.01 to 0.14); continuous measurements also captured within-group longitudinal variation—including cases where fibrosis stage remained unchanged but quantitative values still varied; in a subset of 50 biopsies enriched for NAS grade-2 ballooning, the median distance from ballooned hepatocytes to perisinusoidal fibrosis was significantly shorter than that to lipid droplets (p < 0.0001). Authors’ positioning verbatim: Existing data support its use as an auxiliary quantitative research platform—not as a replacement for routine H&E assessment or independent clinical grading; inflammatory cell outputs failed to reproduce conventional lobular inflammation grading. This article’s assessment: The gap between cohort one’s 0.85–0.88 and cohort two’s 0.50–0.54 suggests the reference standard itself is drifting (the two cohorts differ in methodology: the former compares against the median score of four pathologists; the latter uses exploratory pooling of trial-read slides—this methodological difference is this article’s summary of the abstract); cohort two’s lack of correction for repeated observations means the gap cannot be fully attributed to reference drift; the inflammatory cell failure case demonstrates that deep learning does not automatically reproduce conventional semiquantitative grading—each metric requires individual validation.
Hepatology Communications’ March issue published that follow-up study (teams from Sorbonne University and Pitié-Salpêtrière Hospital, Paris, among others; 10 authors; PubMed XML contained no COI declaration field, so none is reported here): AI-assisted single-fiber quantitative digital pathology measured fibrosis change after bariatric surgery. Fifty-eight patients had preoperative and postoperative (>1 year) liver biopsies, with mean follow-up of 5.84 years (SD 2.59); conventional histology classified 40 as improved, 12 as unchanged, and 6 as progressed (40 + 12 + 6 = 58, recalculated herein). Median continuous phenotypic fibrosis composite score (Ph-FCS) decreased from baseline 5.25 to follow-up 3.75 (p < 0.001); baseline Ph-FCS correlated with postoperative decline (r = 0.72, p < 0.001); declines were −1.78 (95% CI −2.15 to −1.04) for those with histologic improvement ≥1 stage, −1.08 (−1.93 to 0.01) for unchanged, and −0.28 (−1.26 to 0.71) for progressed, with intergroup differences p < 0.02. Two details: Ph-FCS identified intra-stage fibrosis regression in 9 of the 12 patients deemed 'unchanged' by conventional histology; among the 34 patients with baseline advanced-stage fibrosis, 15 remained classified as advanced-stage at follow-up by histologic staging—but their Ph-FCS declined from baseline (though remaining higher than that of the 19 who regressed; 15 + 19 = 34, recalculated herein). Authors’ conclusion: Automation, quantification, and greater sensitivity to change than conventional pathology can improve feasibility of antifibrotic trials. This article’s assessment: The value of continuous readouts lies in rendering 'no stage change but lesion regression' into a detectable signal (9 of 12 'unchanged' cases showed intra-stage regression); context matters: bariatric surgery is a strong intervention with large declines; 'greater sensitivity to change' does not automatically translate into surrogate trial endpoints—trial-level evidence is required, and the authors themselves position this only as 'improving feasibility'.
## Populations Are Changing, Tools Have Not Kept Pace
The Korean Journal of Radiology published a review (Samsung Medical Center, Sungkyunkwan University School of Medicine; two authors; declared no potential conflicts of interest), highlighting a structural problem: Noninvasive imaging diagnosis of HCC currently relies on 'arterial-phase hyperenhancement plus delayed-phase washout', criteria designed for guideline-defined high-risk populations (chronic HBV, chronic HCV, cirrhosis), aiming to achieve simultaneously high positive predictive value and sufficient specificity within clearly defined groups. Yet HCC incidence is rising outside traditional high-risk groups—including among individuals with MASLD and genetic susceptibility—and HCC in non-high-risk individuals is often diagnosed at later stages. The binary 'high-risk/non-high-risk' classification is under pressure. The review proposes integrative directions including imaging features, fibrosis staging, serum biomarkers, individualized risk models, and AI-assisted cross-sectional imaging; all are labeled at the review level as 'worthy of further investigation' as evidence accumulates (this article paraphrases the source wording).
A paper in the Journal of Gastrointestinal Oncology (teams from University Hospital of Cuenca, Spain, among others; four authors; declared no conflicts of interest) has no abstract recorded in PubMed; this article cites only its title and does not paraphrase content: Its title advocates re-evaluating the alpha-fetoprotein (AFP) surveillance threshold in the MASLD era and calls for evidence-based recalibration (title-level citation; this article does not paraphrase the main text).
BMC Gastroenterology published an unsupervised clustering study in multiethnic populations (teams from Ningbo No. 2 Hospital, Northeastern University, Xinjiang Medical University, among others; 14 authors; declared no competing interests): In a cross-sectional analysis of the Urumqi population, MASLD prevalence was significantly higher among Han Chinese than Uyghurs (9,087 cases, 29.81% vs. 15,958 cases, 12.42%, P < .001); among MASLD patients, the proportion classified as high-risk for advanced fibrosis by FIB-4 was 7.70% in Han Chinese versus 2.32% in Uyghurs (P < .001)—more than threefold difference. Patients clustered into five subtypes, characterized primarily by metabolic health, hypertension, dyslipidemia, low LDL-C, and elevated liver enzymes; the dominant features of subtypes were consistent across ethnicities; however, the subtypes associated with more advanced fibrosis differed: low LDL-C in Han Chinese, hypertension in Uyghurs; abnormal glucose metabolism was a consistent feature of high-risk fibrosis subtypes across ethnicities. A retrospective cohort of Han Chinese in Ningbo provided longitudinal follow-up; directional trends in FIB-4 change within subtypes aligned with cross-sectional results. This article’s assessment: Clustering structure is replicable, but 'which subtype carries highest risk' is not generalizable across populations; a single risk model cannot be universally applied; moreover, the FIB-4 high-risk proportion itself differs by over threefold between the two populations—echoing earlier points: identical numerical readouts drift across populations, and any conclusion about 'substituting a benchmark' requires population anchoring.
## Three Assessments
First, 'models outperforming FIB-4' remains unresolved: None of the three lines of evidence delivers a definitive answer. Systematic reviews identified only three head-to-head comparisons—all conducted by original development teams—with combined ΔAUROC confidence intervals crossing zero; CT scoring’s honest direct comparison point estimate lagged by 0.03 (p = 0.37, no statistical significance), and its incremental confidence interval excluded improvements >0.05; a decision-support prototype, after leakage-controlled self-review, achieved 0.81 versus FIB-4’s 0.92 on the same locked test partition. FIB-4’s standing rests not on how accurate it is, but on its availability, interpretability, and status as 'the comparator every new model must face'—raising the burden of proof for challengers across all three evidence types (this article’s assessment).
Second, the most robust advances in this literature occur in 'measurement', not 'diagnostic modeling': The three heaviest-weighted works output no new diagnostic classifier; ROI standardization (inter-physician consistency rate 81%, κ = 0.25, AI–physician consistency 93.8%) improves measurement reproducibility; viscosity exposure (β = 5.80) improves input validity; continuous pathology readouts (9 of 12 'unchanged' cases showing intra-stage regression) improve endpoint resolution; none directly lifts any AUC, yet all determine whether inputs and endpoints for downstream models are trustworthy (this article’s synthesis).
Third, populations and surveillance have not kept pace with shifting disease epidemiology: Rising HCC in non-high-risk populations during the MASLD era—with later diagnosis—calls for AFP threshold recalibration (a title-level signal), differing high-risk fibrosis subtypes across ethnicities, and over-threefold disparity in FIB-4 high-risk proportions collectively point to the same gap: existing noninvasive tools were largely designed and validated in HBV-, HCV-, and cirrhosis-dominant populations; metabolic populations require recalibration—a challenge involving both data and trial design (this article’s synthesis, aligning directionally with the Korean Journal of Radiology review).
## What Different Readers Can Take Away
Hepatologists and gastroenterologists: Initiating first-line stratification with FIB-4 remains reasonable (review conclusion explicitly states 'does not support replacement'); for newly proposed scores claiming superiority (e.g., proteomic or machine-learning combinations), note that comparative evaluations in this literature all originate from original development teams—lacking independent third-party validation; when interpreting liver stiffness values, remember viscosity confounding—especially when comparing timepoints before and after anti-inflammatory therapy.
Radiologists and sonographers: κ = 0.25 is both a warning and a quality-control entry point: ROI standardization has reader-level evidence (93.8% agreement), but has not yet demonstrated improved staging endpoints; MR staging model performance should be read from external test values (F1–3 vs. F4: 0.791; another category: 0.829); dual-site implementation within one institution does not equal cross-institutional generalizability.
Model developers: Internal fitting, locked test partitions, leakage-controlled re-evaluation, reporting 'upper bound of incremental gain relative to FIB-4', and archiving both development- and re-evaluation metrics publicly: CT scoring and HepDecide demonstrate that negative results are equally publishable and citable; the gap between 0.98 and 0.81 serves as a cautionary signal for all 'high development-phase scores'.
Trial designers and regulators: Continuous pathology readouts (Ph-FCS, Tri-MASH) demonstrate enhanced sensitivity to change, but authors self-position them as research platforms or 'feasibility improvers'—not surrogate endpoints; trial screening cohort performance (e.g., resmetirom cohort) cannot be directly extrapolated to real-world clinical practice; studies with deep industry involvement (e.g., Madrigal employee/shareholder authors, GE-provided external validation data, Aiforia platform usage) should be disclosed verbatim and independently re-reviewed.
Patients and the public: These studies do not alter any individual’s lab tests, medications, or surveillance decisions; liver disease screening, elastography, and HCC surveillance plans remain determined individually by patients and their physicians.
Scope and boundaries: All thirteen sources cited herein are abstract-level materials (one lacks an abstract, cited title-only); full texts were not accessed. Dates reflect PubMed-recorded publication or online dates—where one record lists only year, and two list only month; HepDecide’s three archived development-phase values (0.90, 0.98, 0.92) and their corresponding endpoints were inferred by this article based on source-listed order; the source did not annotate endpoints individually, and the original input table and individual test probabilities for the cirrhosis item are no longer retrievable; in the CT scoring additive comparison, 0.80 and 0.79 are source-reported rounded values, and +0.003 is the source-reported difference—this article reproduces these without recalculation; one source abstract reads '0.861and' (missing space), reproduced verbatim as a source typo; MASHPredict’s two external test cohorts and 'Lean' model variable composition were not elaborated in the abstract—this article does not supply them; the multiethnic clustering study employed cross-sectional and retrospective follow-up design—causal inference is not drawn; conflict-of-interest and funding information is reported strictly as declared in each abstract and XML record—this article makes no assertion of 'no conflicts of interest'; three visible industry engagement signals (pharma employee/shareholder authors in trial screening cohort research, device manufacturer employee and dataset source in elastography research, platform vendor employment in pathology model research) are all reproduced verbatim in the main text.
Sources: Cureus, October 2026 (PMID 42936434, https://doi.org/10.7759/cureus.117908); Abdominal Radiology, November 3, 2026 (PMID 42975642); Healthcare, October 12, 2026 (PMID 42940443); Hepatology Communications, November 1, 2026 (PMID 42951905); Frontiers in Immunology, 2026 (PMID 42959101); Abdominal Radiology, November 5, 2026 (PMID 42986432); European Radiology, November 3, 2026 (PMID 42975708); Academic Radiology, October 29, 2026 (PMID 42957469); Virchows Archiv, November 7, 2026 (PMID 42995264); Hepatology Communications, March 1, 2026 (PMID 42993745); Korean Journal of Radiology, November 2026 (PMID 42984198); Journal of Gastrointestinal Oncology, October 30, 2026 (PMID 42996869, title-only); BMC Gastroenterology, August 8, 2026 (PMID 42989132)