Fibrosis Staging (LiFS): Imaging-Based Evaluation

Updated 14 July 2026

Fibrosis Staging (LiFS) is a non-invasive method using imaging data to grade liver fibrosis severity as an alternative to biopsy.
It leverages multi-modal techniques including CT, MRI, and ultrasound to enhance diagnostic accuracy and monitor disease progression.
Research reveals challenges such as data heterogeneity, modality misalignment, and the need for robust, real-world benchmarks.

Fibrosis Staging (LiFS) denotes the estimation of liver fibrosis severity from non-invasive data, most prominently imaging, with recent challenge literature using the acronym specifically for the Liver Fibrosis Staging task in multi-sequence MRI benchmarks (Liu et al., 8 Dec 2025). The clinical motivation is consistent across modalities: fibrosis stage is tightly linked to prognosis, while liver biopsy remains the reference standard yet is invasive, expensive, subject to sampling error, and affected by inter-/intra-observer variability (Jana et al., 2020). In NAFLD, fibrosis stage is described as the strongest predictor of disease-specific mortality, and cirrhosis is the severe scarring of the liver that marks advanced disease (Jana et al., 2020, Zeng et al., 23 Feb 2025). Taken together, the literature suggests a progression from small, single-source proof-of-concept systems toward real-world, multi-center LiFS benchmarks with missing modalities, spatial misalignment, and out-of-distribution evaluation (Liu et al., 25 May 2026).

1. Clinical rationale and scope

LiFS is not a single algorithmic family but a problem setting: infer fibrosis or cirrhosis severity from imaging or imaging-derived measurements, with the aim of supporting diagnosis, triage, and follow-up. The task appears in NAFLD/NASH, chronic hepatitis B, and broader chronic liver disease cohorts, and it spans CT, ultrasound, pathology, multiparametric MRI, elastography-related measurements, and MRI-derived quantitative biomarkers (Jana et al., 2020, Liu et al., 2020, Li et al., 2024).

The role assigned to LiFS varies by study. Some works target direct replacement or partial substitution of biopsy for screening or longitudinal monitoring; others position LiFS as computer-aided diagnosis that augments radiologists or pathologists rather than replacing them (Jana et al., 2020, Liu et al., 25 May 2026). In the CARE-Liver literature, LiFS is framed explicitly as a clinically realistic MRI task in which fibrosis must be inferred from heterogeneous, incompletely available, and often unregistered sequences acquired across centers and vendors (Liu et al., 8 Dec 2025).

A recurring theme is that fibrosis severity is biologically ordered, but the observable imaging phenotype is neither homogeneous nor modality-invariant. This is especially visible in intermediate disease, where moderate-stage cirrhosis/fibrosis is repeatedly described as the hardest class to distinguish (Zeng et al., 23 Feb 2025). The same difficulty appears in binary benchmark decompositions, where robust discrimination of clinically important thresholds such as S4 vs. S1–S3 and S1 vs. S2–S4 remains substantially easier in-distribution than under cross-center shift (Liu et al., 25 May 2026).

2. Label spaces, reference standards, and task formulations

LiFS does not use a universal staging ontology. Different studies adopt different grading systems, then frequently collapse them for learnability or clinical decision support.

Study	Label space	Training/evaluation formulation
CT + pathology (Jana et al., 2020)	Original fibrosis distribution $\{0,1,2,3,3.5,4\}$ , merged to class 0 = stage 0, class 1 = stages 1 and 2, class 2 = stages 3, 3.5, and 4	Multiclass classification
MRI cirrhosis staging (Zeng et al., 23 Feb 2025)	Mild, Moderate, Severe	3-class ordinal classification problem, optimized with standard cross-entropy classification
Ultrasound HBV cohort (Liu et al., 2020)	Normal; Coarseness of liver parenchyma echo; Liver fibrosis; Liver cirrhosis	4-class diagnosis problem
MRI radiomics with biopsy (Huang et al., 9 Sep 2025)	Ishak F0–F6, collapsed to F0–F2 vs. F3–F6	Binary significant fibrosis classification
CARE LiFS / LiQA (Liu et al., 8 Dec 2025)	Scheuer-style $S1$ – $S4$ staging with endpoints $S1\text{--}3$ vs. $S4$ and $S1$ vs. $S2\text{--}4$	Two binary subtasks

Ground truth is likewise heterogeneous. Some systems use pathology from biopsy or surgical resection within three months of MRI, which gives strong clinical validity to labels in the LiQA and LiFS benchmarks (Liu et al., 8 Dec 2025, Liu et al., 25 May 2026). Earlier CT-pathology fusion work obtains fibrosis stage and NAS from a pathologist’s manual examination of whole-slide images, then predicts those labels from CT or CT+pathology (Jana et al., 2020). The MRI radiomics study uses Ishak scoring by two experienced pathologists blinded to MRI and clinical data (Huang et al., 9 Sep 2025). Ultrasound studies often rely on experienced radiologists’ diagnostic labels or routine clinical fibrosis assessments rather than explicit histologic stage per image (Liu et al., 2020, Li et al., 2020).

A notable methodological tension is that many studies acknowledge ordered severity while still training standard classifiers. The CT study explicitly does not frame fibrosis as ordinal regression (Jana et al., 2020), and the MRI cirrhosis study states that the task is clinically an ordinal progression problem but again uses standard cross-entropy classification rather than an ordinal loss (Zeng et al., 23 Feb 2025). Later benchmark work reintroduces ordinal modeling more directly, for example through soft-label ordinal regression heads in LiQA (Liu et al., 8 Dec 2025).

3. Data modalities and input representations

CT-based LiFS commonly reduces 3D volumes to instance collections. In the 30-patient NAFLD dataset, each CT scan is split into 2D slices, windowed to $[-200,250]$ , segmented so that non-liver pixels are removed, and filtered to discard slices with too little liver; the resulting cohort contains 2,595 CT slices (Jana et al., 2020). When pathology is available, H&E whole-slide images are tiled into non-overlapping $224 \times 224$ patches at $5\times$ magnification after background removal, yielding 7,775 pathology patches (Jana et al., 2020). A related CT-only line uses the same small 30-patient setting but adds self-supervised pretraining and texture-encoded inputs via Local Binary Pattern (LBP) to emphasize subtle parenchymal texture changes associated with fibrosis (Jana et al., 2021).

Ultrasound LiFS is highly view-dependent. One framework uses 10 indicator ultrasound images per patient, 13 indicator labels, and a final 4-class diagnosis label, explicitly mirroring a radiologist workflow in which multiple views support intermediate findings and the final fibrosis decision (Liu et al., 2020). Another study-level ultrasound system processes up to 14 ultrasound images in a study, restricts attention to a clinically defined ROI containing the liver parenchyma and upper border, and fuses an arbitrary number of image features through global hetero-image fusion (GHIF) (Li et al., 2020). A separate hybrid system combines ultrasound images with blood-test probabilities and outputs the three classes No Fibrosis, Fibrosis, and Cirrhosis (Kashyap et al., 28 Apr 2025).

MRI LiFS ranges from relatively clean single-sequence classification to real-world multiparametric staging. Cirrhosis stage estimation has been studied on separate T1W and T2W datasets, with sequence-specific encoders and attention guidance (Zeng et al., 23 Feb 2025). In LiQA and later LiFS benchmarks, MRI becomes multi-center, multi-vendor, and multi-sequence, typically including T1W, T2W, DWI, and four-phase gadoxetic acid-enhanced imaging; the data are explicitly described as having missing modalities and no spatial registration (Liu et al., 8 Dec 2025, Liu et al., 25 May 2026).

The representational space is broader than end-to-end CNN inputs. One CARE 2025 pipeline derives 32 STAD features—Shape, Textural, Appearance, and Directional—from segmented liver ROIs and classifies fibrosis with Random Forests (Zhou et al., 30 Sep 2025). Another MRI study uses PyRadiomics features from T2WI, CSE, and diffusion-derived parametric maps for binary significant fibrosis classification (Huang et al., 9 Sep 2025). At the mechanistic end, an ex-vivo MRI study models susceptibility as $S1$ 0, interpreting the negative component as fibrosis-related diamagnetic source signal (Li et al., 2024).

4. Core modeling paradigms

A first paradigm is instance aggregation. The CT/pathology work uses a ResNet-18 feature extractor without the final fully connected layers, followed by fully connected layers with 512 and 128 neurons, average pooling across slice- or patch-level features, and a final patient-level classifier (Jana et al., 2020). When both CT and pathology are present, two fusion strategies are compared: mid-fusion, which concatenates local features before the classifier, and late-fusion, which concatenates global pooled features. The best fibrosis result in that study is mid-fusion, single-loss, using $S1$ 1 rather than $S1$ 2 (Jana et al., 2020).

A second paradigm is representation learning under data scarcity. Self-supervised CT LiFS uses context restoration pretraining on LiTS CT volumes, then fine-tunes a lightweight classifier on the 30-patient fibrosis cohort. The system combines self-supervised learning, adversarial loss during pretraining, and LBP texture encoding during downstream classification, reflecting the hypothesis that fibrosis is expressed through subtle local texture rather than large semantic lesions (Jana et al., 2021). In ultrasound, the analogous small-data response is indicator-guided multi-task learning: multiple shared-weight VGG-16 streams predict both intermediate ultrasound indicators and the final fibrosis diagnosis, trained through a weighted sum of cross-entropy losses and a 3-stage optimization strategy (Liu et al., 2020).

A third paradigm is multi-view and uncertainty-aware fusion. GHIF with view-specific parameterization tailors shared ultrasound CNN processing to different liver viewpoints by altering only normalization parameters, then fuses per-image ROI-pooled features using mean, variance, and max statistics (Li et al., 2020). MRI multi-view systems go further by treating local liver patches and the global ROI as separate views. One framework maps per-view evidence into subjective-logic opinions and combines them through Dempster-Shafer evidence theory (Gao et al., 2023). MERIT extends this design with a distribution-aware base rate, cumulative belief fusion for local views, and belief constraint fusion between local and global opinions, with uncertainty mass used both for reliability and interpretability (Liu et al., 2024).

A fourth paradigm is segmentation-conditioned LiFS. In LiQA, staging is a three-step pipeline: liver VOI extraction, per-view fibrosis scoring using a 3D ResNet-34, and multi-view consensus fusion by averaging predictions across available sequences; CAM-SEM Loss constrains Class Activation Maps to the liver region (Liu et al., 8 Dec 2025). Another CARE method uses segmentation masks to crop liver ROIs, extracts STAD features, adds a vendor flag, and trains Random Forests (Zhou et al., 30 Sep 2025). A third uses LiSeg masks and rigidly aligned multi-parametric MRI to extract overlapping $S1$ 3 liver patches, trains a ResNet-18 + MLP on Stage 1 versus Stage 4 patches only, and maps the proportion of Stage-4-like patches into subject-level probabilities through thresholded piecewise functions (Wang et al., 10 Feb 2026).

A fifth paradigm is missing-modality compensation. Improved mmFormer retains modality-specific and modality-correlated encoders, introduces a missing-modality compensation module with zero-padding, modality-availability masks, and a Delta Function with learnable statistical parameters, then stabilizes inference with a four-model cross-validation ensemble and soft voting (Zhang et al., 19 Sep 2025). This design addresses a central real-world LiFS constraint: arbitrary combinations of absent MRI sequences.

5. Benchmarks and reported performance

Reported performance is highly dependent on cohort size, label granularity, and whether evaluation is internal, cross-validated, or truly out-of-distribution.

Setting	Evaluation protocol	Reported result
CT / pathology fusion (Jana et al., 2020)	3-fold cross-validation; multiclass fibrosis AUC with 95% confidence intervals from 1,000 bootstrap iterations	CT only: $S1$ 4; pathology only: $S1$ 5; mid-fusion, single-loss: $S1$ 6
CT self-supervision (Jana et al., 2021)	3-fold cross validation, averaged over 5 repeated experiments	Fibrosis AUC $S1$ 7
Ultrasound GHIF + VSP (Li et al., 2020)	5-fold cross validation at patient level	partial AUC $S1$ 8, AUC $S1$ 9, R@P90 $S4$ 0
MRI cirrhosis staging (Zeng et al., 23 Feb 2025)	Train/validation/test split on CirrMRI600+	Best T1W model: $S4$ 1 accuracy; best T2W model: $S4$ 2 accuracy
Real-world LiFS benchmark (Liu et al., 25 May 2026)	ID/OOD evaluation with radiologist comparison	Best AI in contrast-enhanced ID: AUC $S4$ 3 for $S4$ 4 vs. $S4$ 5, AUC $S4$ 6 for $S4$ 7 vs. $S4$ 8

Several patterns recur. On small internal datasets, multimodal fusion and carefully structured ROI or view modeling can yield strong cross-validated performance. In the 30-patient CT/H&E cohort, fused CT-pathology outperformed both unimodal baselines, which the authors interpret as complementary information between global morphologic CT cues and fine tissue-level histology (Jana et al., 2020). In ultrasound, both indicator guidance and multi-image study-level fusion materially improve over simpler baselines (Liu et al., 2020, Li et al., 2020).

MRI radiomics remains competitive in binary staging. A biopsy-linked study reported a best fibrosis combination of {T2WI, CSE, Diffusion} with test AUC 0.89, while an integrated model that included the inflammation prediction score reached AUC test 0.97 (Huang et al., 9 Sep 2025). The ex-vivo diamagnetic susceptibility study reported AUC = 0.987 for $S4$ 9 vs. $S1\text{--}3$ 0 using $S1\text{--}3$ 1-based negative susceptibility $S1\text{--}3$ 2, although that result is explicitly feasibility-stage and ex vivo (Li et al., 2024).

The benchmark literature is more conservative. In LiQA, validation AUCs peaked at 0.8108 for cirrhosis and 0.8056 for substantial fibrosis under ordinal regression, yet the final CE wCSL submission reached only 0.6481 \pm 0.1072 and 0.5931 \pm 0.2084 AUC on the test set (Liu et al., 8 Dec 2025). On the larger 610-patient LiFS benchmark, the best AI systems were broadly comparable to the senior radiologist in selected in-distribution cells and significantly exceeded the junior radiologist in selected settings, but median AI performance degraded sharply under out-of-distribution shift (Liu et al., 25 May 2026). This suggests that apparent maturity in curated cohorts does not automatically transfer to heterogeneous clinical deployment.

6. Reliability, interpretability, and unresolved issues

The main technical obstacles are now clearer than the headline accuracies. Cross-center heterogeneity, label imbalance, missing modalities, and contrast-enhanced sequence variability, especially in HBP, are identified as dominant benchmark-level challenges (Liu et al., 25 May 2026). One important consequence is metric selection: the LiFS benchmark explicitly shows recurring values such as 92.86% ACC paired with low AUCs, making clear that accuracy can be misleading under strong imbalance and that AUC is often the more reliable comparator (Liu et al., 25 May 2026).

Interpretability is pursued through several distinct mechanisms. Ultrasound indicator-guided learning uses clinically meaningful intermediate labels (Liu et al., 2020). CAM-based regularization in LiQA constrains attention to liver anatomy (Liu et al., 8 Dec 2025). Patch-based MRI LiFS visualizes red and blue Stage-4-like versus Stage-1-like regions across the liver (Wang et al., 10 Feb 2026). STAD-based pipelines are explicitly tied to measurable abnormalities such as surface nodularity, parenchymal heterogeneity, and directional tissue organization (Zhou et al., 30 Sep 2025). Evidential multi-view methods expose both per-view belief and uncertainty, and Dempster-Shafer-type fusion rules make the combination logic inspectable rather than purely latent (Gao et al., 2023, Liu et al., 2024).

Several controversies are methodological rather than clinical. One is whether LiFS should be trained as conventional classification or as an ordinal problem. Another is how much the field should rely on segmentation-first pipelines, given that segmentation errors, registration inaccuracies, and pseudo-label noise propagate into staging (Zhou et al., 30 Sep 2025, Wang et al., 10 Feb 2026). A third is modality dependence: T1W often outperforms T2W in MRI cirrhosis staging (Zeng et al., 23 Feb 2025), while missing-modality compensation methods indicate that robustness to incomplete acquisitions may matter as much as nominal in-distribution accuracy (Zhang et al., 19 Sep 2025).

There are also modality-specific confounders. In elastography-related measurements, fat and fibrosis jointly influence both shear wave speed and shear wave attenuation; at a constant level of fat, fibrosis stages can influence SWA by factors of 2–4, and the SWA-fat relationship depends strongly on fibrosis stage (Poul et al., 2020). This directly cautions against one-parameter staging heuristics in steatotic liver disease. Conversely, susceptibility-based MRI suggests a more mechanistic fibrosis biomarker by targeting the diamagnetic contribution of collagen-rich scar tissue, but the present evidence is still ex vivo and small-scale (Li et al., 2024).

A plausible implication is that LiFS is converging toward a composite design space rather than a single dominant architecture: anatomically constrained inputs, modality-aware or missing-modality-aware fusion, explicit uncertainty handling, and evaluation under real domain shift. The benchmark results indicate that top systems can approach senior-radiologist performance in selected settings, but the same literature also shows that clinically reliable deployment still depends on solving robustness, calibration, and intermediate-stage ambiguity rather than only improving in-cohort accuracy (Liu et al., 25 May 2026).