ABSTRACT
Metabolic dysfunction associated steatotic liver disease is the most common chronic liver disease worldwide, associated with considerable clinical and economic burden. Early detection and timely intervention of metabolic dysfunction associated steatohepatitis (MASH) remains a key cornerstone of management. Traditionally, histopathology assessment is central to the characterization of MASH disease activity and fibrosis, albeit with inherent limitations. The advent of artificial intelligence (AI) has provided transformative tools to augment many aspects of MASH diagnostics, including histology assessment. New technologies enabling digitalization of pathology slides and allowing for further AI assisted model analysis have helped to address some of the previous limitations. In this review, we highlight the key AI assisted digital pathology tools, including utility, performance and clinical impact. Limitations and real-world considerations in adopting these AI tools are also explored.
-
Keywords: Artificial intelligence; Digital pathology; Metabolic dysfunction steatotic liver disease
INTRODUCTION
Metabolic associated steatotic liver disease (MASLD) is the most common chronic liver disease, and is estimated to affect more than 1 billion adults worldwide [
1]. Metabolic dysfunction-associated steatohepatitis (MASH), the progressive form of MASLD, can lead to cirrhosis, hepatocellular carcinoma (HCC), and liver events including mortality. MASLD is also associated with extrahepatic complications, such as chronic kidney disease, cardiovascular disease, and cancer [
2]. Mirroring the obesity epidemic, the prevalence of MASLD, MASH and potential complications is anticipated to increase, representing a significant clinical and economic burden. The advent of artificial intelligence (AI) has permeated all facets of living and promise to reduce workload burdens by automating routine repetitive tasks that may be time consuming, or augment focus in other areas such as data analytics and interpretation, all of which are pertinent to the challenges of addressing the MASLD epidemic. One major area AI has made inroads in is assisting digital pathology (DP) imaging analysis in MASH histology assessment.
The aim of this review is to recapitulate the role of histopathology and DP in MASH, and to explain the main AI models that have been developed for MASH diagnostics. Prior reviews have covered extensively the technical aspects of AI applications in MASH [
3,
4]. We provide a plain and readable overview and update of the main applications and potential clinical value of AI technologies and cover important considerations in their real world use.
ROLE OF HISTOPATHOLOGY IN MASH
Histopathological assessment of a liver biopsy, despite its invasive nature, remains the gold standard for the diagnosis of MASH. It is also the reference standard for evaluation of MASH severity and assessment of treatment response. The two major components to be assessed are the grading of the disease activity, and staging of fibrosis. These are performed using validated assessment tools by pathologists, such as the widely used non-alcoholic fatty liver disease activity score (NAS) by the non-alcoholic steatohepatitis clinical research network (NASH-CRN) [
5]. Other methods, such as SAF (steatosis, activity, fibrosis) by Bedossa et al. [
6] have also been described. The degree of liver fibrosis histologically is able to prognosticate the progression of MASH and development of clinically relevant outcomes [
7,
8].
Histopathology is especially relevant in MASH clinical trials, as short-term evidence of efficacy. Given the long latency associated with clinical manifestations of liver-related events, clinical outcomes as endpoints are not realistic in Phase I-II studies because of small sample size and short duration of follow-up. In Phase III studies, while waiting for clinical outcomes to manifest, conducting interim analysis based on histological response can facilitate accelerated drug approval. Trials recruitment has also traditionally relied on histopathologic classification to determine eligibility for inclusion. Standardised grading and staging systems, such as the ones devised by NASH-CRN, allow for the semi-quantification of steatosis, inflammation, hepatocellular ballooning and fibrosis. Current clinical trial endpoints adopted by agencies such as the Food and Drug Administration (FDA) include resolution of steatohepatitis and no worsening of liver fibrosis, improvement in liver fibrosis without worsening of steatohepatitis, and both resolution of steatohepatitis and improvement in fibrosis [
9]. The value of the liver biopsy is well established, providing a level of confidence in the diagnosis, being able to exclude potential secondary concurrent chronic liver disease or even discovering histologic/genomic phenotypes that predict drug response. Although non-invasive assessments of MASH and liver fibrosis have recently been recognised by the FDA as reasonably likely surrogate endpoints, it remains to be clarified how this impacts on the rigor of enrolment, as well as treatment response and real world clinical care by extrapolation. Significant discordance can exist between non-invasive fibrosis assessment and liver histopathology [
10]. As such, liver biopsy with histopathologic classification and assessment of severity remains relevant for MASH trials recruitment, and assessment of treatment response.
While histopathologic assessment of MASH is the reference standard in clinical trials, it does have some notable shortcomings. One major drawback of the current scoring systems is intra-and inter-observer variability [
11-
15]. A study of 678 liver biopsies from the EMMINENCE trial showed not just poor inter-rater reliability of NAS scoring, but also real world implications; 46% of patients had disagreements on whether they met the qualifying criteria for the study; for outcome assessment there was disagreement of 19–23% for evaluation of NASH resolution without worsening fibrosis, and 21–24% for evaluation of fibrosis improvement without worsening NASH [
16]. The use of conventional ordinal scoring systems also reduces granularity, and therefore sensitivity to small changes in disease state. This has implications for MASH trials recruitment, which uses histopathology to assess eligibility for entry into the trial. It also impacts the final determination of trial outcomes, given that it is used as the primary surrogate outcome in most registration trials to date.
In view of the above-mentioned limitations of current histopathology methods to assess MASH and liver fibrosis, there is considerable interest in the use of DP and AI to improve these systems for risk stratification, diagnosis, monitoring, and prognostication of MASH [
17]. The digitalization of pathology has opened up multiple avenues to address these issues. We elaborate on these below.
DIGITAL PATHOLOGY FOR MASH
Digitalization of conventional glass slides has afforded new opportunities to address the limitations of traditional histopathology in MASH and improve workflows, reducing cost. First, DP can facilitate clinical workflows such as central reading and slide image sharing, thereby allowing speedy consensus in diagnosis and scoring [
18,
19]. Second, DP images can be further analysed by image processing algorithms, including AI models. These can also be used to optimise reproducibility of histologic grading. Third, new methods of image acquisition from glass slides can extract signals not available from conventional light microscopy. One new technology of relevance to MASH is second harmonic generation/two photon excitation fluorescence (SHG/TPEF) [
20] microscopy, which allows acquisition of signals not available from conventional microscopy (e.g., detection of collagen fibres with SHG, and hepatocytes with TPEF). Current commercial platforms may combine multiple of these methods - for example, FibroSight by HistoIndex combines SHG/TPEF on the Genesis platform with further AI analysis of the images.
Digital image processing algorithms for MASH have made significant progress in recent years. Initial efforts used traditional image processing techniques, such as thresholding of colour and hue parameters to identify collagen in Sirius Red stained slides. Subsequently, AI models have been trained to identify other specific features in histology slides, such as collagen and lipid droplets. As AI models are able to perform the segmentation across the entire slide, whole slide quantitative measures can be performed, a task that conversely cannot be performed easily by a pathologist. An example of this is the collagen proportional area, where the relative amount of collagen in the entire slide is calculated. Both traditional image thresholding techniques [
21] and deep learning models [
22] have been used to perform this task. Overall, harnessing DP techniques allows a more consistent and precise quantification of MASH parameters than the semi-quantitative staging performed by humans, and have been extended to allow whole slide scoring of MASH components (i.e., steatosis, lobular inflammation, ballooning, and fibrosis).
We elaborate below on the state of the art in AI techniques. Two contemporary platforms that we discuss are FibroSight by HistoIndex [
23], and AIM-NASH by PathAI [
24]. AIM-NASH has recently been qualified by the FDA and the European Medicines Agency (EMA) for use in clinical trials. A third platform is FibroNest by PharmaNest [
25]. We summarise the salient features of these technologies and their relevance to MASH.
FibroSight
The FibroSight platform aims to reduce both analytical variability, and pre-analytical errors from staining, by optimising both stages of the pathology reading process. SHG/TPEF is first used to acquire digital images of the slide, which can then be digitally analysed for fibrosis and steatosis related parameters, using an AI model.
SHG/TPEF uses a femtosecond laser to interrogate unstained slides. This is a form of nonlinear optical microscopy that produces label-free images with sub-micron spatial resolution. Photons emitted from the sample form the SHG signal, which detects and quantifies fibrillar collagen in two-dimensional optical sections of the tissue. Other photons emitted from the sample form the TPEF signal, which allows visualization of the background hepatocytes and lobular architecture. As no staining is required and scanning is standardised, this approach reduces pre-analytical variability associated with staining and slide-scanning [
26].
On the HistoIndex Genesis platform, images are acquired with a 780 nm femtosecond excitation source at 20× magnification, yielding a pixel resolution of 0.3906 μm across a 200×200 μm field of view. The intrinsic optical sectioning of two-photon excitation provides uniform signal generation through the full thickness of standard 4–5 μm FFPE sections, and near-infrared excitation enables simultaneous, label-free detection of fibrillar collagen (SHG at 390 nm) and hepatocellular autofluorescence (TPEF at 550 nm) [
20,
23].
The AI model in the FibroSight platform then performs key image analysis steps, including segmentation of histopathologic features, measurements of various collagen morphological parameters to quantify fibrosis, and quantitative scoring of NAS score components. The FibroSight AI model is the culmination of prior efforts to model fibrosis related parameters using SHG. Automated identification of fibrosis parameters using SHG was first described in chronic hepatitis B by Xu et al. [
27] in 2015, and termed the qFibrosis model. This was then applied to MASH by Wang et al. [
28] in 2017 where multiple qFibrosis parameters were investigated. Various fibrosis related parameters were examined, of which the most important were number of collagen strands, strand length, strand eccentricity, and strand solidity. This work was further expanded upon by Liu et al. [
26] in 2020, where the qFIBS model was introduced. This model was able to score not just fibrosis (qFibrosis), but also the other components of the NAS score (qInflammation, qBallooning, and qSteatosis). Further refinements were made by Leow et al. [
29] in 2020 to better differentiate F1 and F2 fibrosis. Since then, the model has also been improved on paediatric MASH patients [
30].
qFibrosis measurements have been shown to be repeatable and reproducible [
31]. When combined with pathologist assessment, Soon et al. [
32] also demonstrated that the use of AI significantly improved the agreement between pathologists when evaluating fibrosis. The benefits of AIassistance with qFibrosis in inter-rater concordance have been further validated using other retrospective cohorts [
33]. This demonstrates the improvements AI-integrated DP affords over manual assessment alone for staging fibrosis in NASH.
AIM-NASH
In contrast to FibroSight, the AIM-NASH (artificial intelligence-based measurement of non-alcoholic steatohepatitis histology) platform is based on analysis of scanned whole slide images. The slides are conventional glass slides stained using H&E and Masson’s trichrome [
19].
The AI model involves two main parts. First, a segmentation model to analyse whole slide images, to identify and segment MASH related features. On H&E, these include steatosis, lobular inflammation, and hepatocellular ballooning. On trichrome, this includes pathological fibrosis. Non-pathological fibrosis subtypes are recognised and excluded by the AI model. The segmentations are presented as digital overlays, which can be reviewed by a pathologist. Second, a graph neural network then makes use of the overlays to predict the overall NAS and CRN fibrosis scores. In its intended use, a pathologist reviews the digital slide, the overlays, and the overall AI model output, before approval of the final scoring [
34].
In its initial development, the AIM-NASH model had better consensus than pathologists. When used as part of the overall platform, there was a high reported concordance with glass reads (i.e., slides that are not digitalised), and the reads were both repeatable and reproducible. Compared to individual manual reads, AIM-NASH had non-inferior accuracy for the four main features of steatosis, lobular inflammation, hepatocellular ballooning, and fibrosis. Similar results were also observed for AIM-NASH assisted pathologist reading. For hepatocyte ballooning, AIM-NASH alone and in assisted reads had superior accuracy to individual manual reads [
34].
The main advantage of AI-assisted reading with models such as AIM-NASH derive from its analysis of the whole slide (which is impossible for pathologists performing manual reading) and the consistency in its reading. In particular the superior performance on scoring of hepatocyte ballooning is encouraging, as human scoring of this item has high variability [
35]. The extensive validation of AIM-MASH in various cohorts, totalling >13,000 patients, has led to its qualification by the FDA [
36] and EMA [
37] as a tool for outcome measurement in MASH clinical trials.
The segmentation overlays generated by the model can also be used for quantitative scoring. This would be a continuous measurement of histologic parameters such as fibrosis, as opposed to the conventional ordinal scale. Although such quantitative parameters have yet to be endorsed as endpoints by regulatory authorities, they do have correlation with clinically relevant endpoints, and may be potential new surrogate endpoints in the future. The original research describing AIM-NASH showed that continuous scoring of fibrosis was associated with progression to cirrhosis and predicting liver related events. Continuous fibrosis scoring also showed more response in treatment responders, as compared to placebo, in a retrospective analysis of the ATLAS trial. AI-assisted continuous scoring can thus detect subtle improvements in MASH that may not amount to one stage in the current scoring system [
24]. It may also be useful as a biomarker for future risk, thereby adding more prognostic value to the liver biopsy.
FibroNest
FibroNest, like AIM-MASH, utilizes digitized whole slide images. The slides are conventional glass slides stained using Sirius red. The platform is able to perform automated NAS scoring. One key distinguishing feature of FibroNest is AI identification of collagen fibers, and various fibrosis phenotypes [
38]. These fibrotypes have been explored as potential markers of molecular alterations in MASH - for example, using the FibroNest algorithm, Fujiwara et al. [
39] showed that distinct fibrosis-related profiles could be identified and that one fibrotype was predictive of incident HCC. On further exploration using spatial single-cell transcriptomic analysis, the fibrotype was correlated with the presence of HCC-promoting hepatic stellate cells.
CLINICAL IMPACT OF AI AND DIGITAL PATHOLOGY IN MASH REAL WORLD DATA
Clinical studies have begun to use DP and AI methods in secondary analyses. The ICONA study [
40], a Phase IIb RCT of the FFAR1/FFAR4 agonist icosabutate, histological improvement assessed by AIM-MASH corroborated the findings of conventional pathology. Quantitative fibrosis scoring using both AIM-MASH and SHG/TPEF (qFibrosis) showed response to the study drug.
Secondary analysis of the MAESTRO-NASH Phase III trial of resmetirom [
41] using qFibrosis identified more individuals with fibrosis response, as compared to central reading by pathologists [
42]. Of note, 20% of baseline patients were scored as F4 by qFibrosis, which in theory would have led to trial exclusion. Admittedly, this discrepancy may be contributed by the use of NAS which provides assessment defined by static categories designed to stage “untreated cases” and not post-treatment staging. This by itself represents an inherent limitation of capturing a dynamic process.
Other retrospective studies of clinical trials data have also showed responsiveness of quantitative fibrosis measures in patients treated with the study drug, including trials which might be deemed negative by conventional histopathology. This includes the FLIGHT-FXR Phase II RCT of the FXR agonist tropifexor [
43], where most patients did not show fibrosis regression by conventional histopathology. However, post-hoc analyses with qFibrosis identified more responders to the study drug. A further re-analysis [
44] showed that subtle regressive changes in septa morphology and collagen fibre parameters were observed using SHG/TPEF in patients with F3 fibrosis, who would otherwise have been deemed to have no response with conventional scoring. As mentioned above, a re-analysis of the ATLAS trial [
24] showed more response of continuous fibrosis score in drug responders. Nevertheless, it remains to be seen whether the subtle changes afforded by these new technologies would translate into meaningful improvements in clinical outcomes. The higher granularity of AI DP may make an otherwise negative study have a small positive histological benefit, only for the drug to fail in trials using clinical outcomes.
Outside of clinical trials, DP is also helpful in determining subtle changes in disease state. In a retrospective study of patients with lifestyle interventions for MASH [
45], SHG/TPEF (qFibrosis) was able to detect regression of fibrosis in the periportal region of patients with strengthened lifestyle interventions, as opposed to those with regular lifestyle interventions. This effect could not be detected on conventional histopathology. Detection of such small changes in DP may open up pathways to grading disease response and better understanding disease mechanisms. However, at present, the clinical significance of such changes has not yet been elucidated.
Beyond quantification of fibrosis and NAS scoring, AI and DP may offer other insights into the broader picture of liver disease. Conventional histopathology staging has prognostic value, and further enhancement with markers derived from DP may enhance predictive value for clinical outcomes. The PathAI team also developed an AI model [
46] which analyses histopathology from patients with MASH compensated cirrhosis, with the aim of predicting hepatic venous pressure gradient (HVPG). The model was able to distinguish patients with normal HVPG, mildly elevated HVPG, and those with clinically significant portal hypertension. qFibrosis has also been used as part of a multivariable risk model to predict adverse liver outcomes in cirrhosis [
47].
We summarise potential DP workflows incorporating AI assistance in
Figure 1. The main steps of a potential sequence would include standard pre-processing workflows, with slides passed along for H&E with digitalization, or other processes required for AI assistance. The pathologist would review the combined final output on a digital workstation, and general a report combining the AI read with their professional interpretation.
CLINICAL LIMITATIONS OF AI AND DIGITAL PATHOLOGY
While we have covered the major advantages of AI-assisted DP, these have mainly been borne out in the setting of clinical trials. It is as yet not determined whether the additional information conferred by these techniques will inform the management of routinely managed MASH patients. While accurate and reliable quantification of the NAS score components is desirable for clinical trials, it has not been shown that these components have long term prognostic value for clinical endpoints. Furthermore, although quantitative fibrosis measurements do have clinical prognostic value, there are at present no population based comparisons against other non-invasive fibrosis assessments, such as elastographic techniques.
Limitations inherent to liver biopsy sampling technique cannot be addressed by AI models at this present time. Geographical variation of MASH within the liver, non-representative or inadequate sampling, and artefacts introduced by specimen preparation cannot be corrected for by AI models. A study examining whether virtually truncated liver biopsy images could still be graded correctly by FibroSight qFibrosis found that results remained reliable and concordant only with biopsy length >20 mm (approximately 10 portal tracts) [
48]. This recapitulates standard requirements for liver biopsy length and stresses the importance of adequate tissue acquisition, even when AI models are used.
Although AI models may be able to review a slide faster than a pathologist, expectations of great efficiency gains may not be easily realised in the short term. Implementation of a DP service adds on additional preprocessing steps, such as slide scanning. Modalities such as SHG/TPEF require an unstained slide, and so cannot be performed in parallel with conventional H&E staining workflows. The use of specific machines for slide scanning (or laser for SHG/TPEF reads) may result in bottlenecks if case volume is high. As the current standard of care requires a pathologist to perform manual slide review, implementation of AI reads may in the short term require more work for pathologists that have to review an additional finding. However, if such technologies are eventually widely adopted, they may become useful co-pilots for rapid turnover of pathology reporting. Finally, investment in DP equipment or services may be substantial, and can be challenging to justify for laboratories outside academic medical centres with lower case volumes.
Present AI models are also not designed to diagnostically distinguish MASH from other histopathologic diagnoses (such as AIH), and neither have they been validated for diagnosis of MASH co-existing with other diagnoses. In the setting of clinical trials for MASH, most patients have generally received thorough clinical screening to exclude other causes of liver disease. However, pathologists using AI-assisted DP must be aware that the systems are currently designed for grading, and not diagnosis
per se. In the context of clinical trials, other diseases that may mimic MASH histopathologically (e.g., alcohol associated liver disease, Wilson’s disease) have generally been clinically excluded. Outside of the trials setting, pathologists must remain attentive to the clinical context in which a liver biopsy was ordered. Finally, the present AI models have mainly been evaluated in adult MASH. Some studies have been performed in other populations - qFibrosis has been extended to paediatric MASH [
30], and it was also originally studied in chronic hepatitis B [
49]. However, the literature on non-MASH populations is less extensive, and adult patients with mixed etiologies (such as MetALD, and concomitant viral hepatitis with MASLD) are not yet well studied. Established interpretation criteria for diseases other than MASH are not well developed. There has been limited scalability of the current AI tools for “non-MASH” liver diseases. In an ideal world, AI should be capable of recognising a range of pathologies. However, this would require additional training, especially for rarer liver diseases. That said, many aetiologies can already be diagnosed through blood tests and standard pathological evaluation, so this is not the most pressing clinical need.
REAL WORLD CONSIDERATIONS IN ADOPTION OF AI AND DIGITAL PATHOLOGY
Besides the clinical benefit and limitations of AI models described above, there also exist other real world considerations to their implementation by clinicians.
One major barrier to adoption of AI by pathologists is the trust required in the results, and acceptance of potential lack of explainability. It would be impossible for a pathologist to manually verify any quantitative fibrosis parameter, and the algorithmic process is inherently a black box (i.e., it is not possible to directly determine how a parameter was measured by the AI model). Data visualisation tools such as slide overlays may allow semi-quantitative verification of the AI output, but an inherent measure of trust by the pathologist is required. Pathologists must recognise that AI output is not intended to replace their clinical judgement, but to function as a helpful assistant. The incorporation of AI tools may accentuate a change in vision to that of a co-pilot concept. In this model, AI functions as an interactive agent rather than a silent automation tool. The AI performs high-volume quantitative tasks (e.g., mitosis counting or fibrosis grading) and alerts the pathologist to subtle regions of interest, while the human expert provides the high-level contextual and clinical integration. This evolution redefines the pathologist as a ‘diagnostic pilot’ who directs the AI’s focus, thereby mitigating transparency issues through constant human-in-the-loop verification and interactive validation of AI-generated insights. Embracing this paradigm can lead to confident and conscious diagnoses. Indeed, a real world study of pathologists scoring MASH cases with the assistance of the HistoIndex platform [
50] showed that use of qFibrosis and visualisation of fine fibrosis features improved concordance among pathologists. AI assistance was rated as useful by >50% of pathologists in the study.
Other aspects of trust include validation of AI models on the pathologist’s local data, as published AI results may not always generalise to new cohorts, outside of curated research samples. Additional features that increase the transparency of how AI algorithms work such as tissue feature visualisation, specificity and sensitivity analysis, uncertainty quantification and bias detection are crucial in explaining AI. We believe that embedding such features into AI models will accelerate adoption of such technology into clinical laboratories.
Another major barrier to the adoption of AI goes beyond individual pathologists, to the system as a whole. Including AI and DP requires changes to laboratory workflow, and upgrading hardware and software components (such as whole slide scanners, high resolution reporting monitors, and servers to store and manage data), and interfacing them to current laboratory information systems and cybersecurity systems. This would require efforts beyond pathologists to all laboratory staff including technical support staff. Support from the wider healthcare system in terms of software support and integration may also be required. With this, cost considerations and accessibility to these services contribute additional complexity to the adoption. Reimbursement for this service can be challenging as it is not currently part of the standard of care, and this may require centralisation of DP services at a larger center for financial viability, if not housed at each individual hospital that equips themselves with these tools. It also remains an open question as to how use of AI should be disclosed in pathology reports, and how liability for results (including any inaccurate results) is assigned. Beyond each healthcare system, appropriate policies need to be in place to guide consistent and responsible use of AI by clinicians in this context. At present, there is no widely used guidance available from professional societies or healthcare regulators on AI use in DP analysis for MASH assessment.
Nevertheless, when DP and AI are implemented, they can still deliver measurable improvements in laboratory efficiency and economic sustainability. The experience of large scale institutions has shown that full digitization can allow time savings of approximately 19 working hours per day, or roughly five minutes per case [
51]. Reported reductions in mean time-to-diagnosis have been reported at over 80 seconds per case, with an overall decrease in turnaround time by up to one full day for complex surgical resections [
52]. Economically, the transition supports long-term fiscal viability; five-year projections suggest potential savings exceeding $1.3 million through reduced manual slide handling, optimized archival access, and the elimination of physical transport for consultations [
52,
53].
We summarize the potential limitations and solutions in
Table 1.
WAY FORWARD IN ADOPTION OF AI AND DIGITAL PATHOLOGY
Advances in technology must be thus matched by transformation at the individual physician level. Willingness to adopt AI and DP workflows, and to master their best use is imperative. As repetitive and quantitative tasks are assigned to AI, pathologists should regard these as any other laboratory tool, and regularly audit the algorithms for accuracy, reliability and reproducibility.
As AI approaches advance, we look forward to new possibilities in AI diagnostics. Integration of accompanying clinical, genomic and non-invasive test data can allow further development of prediction/precision medicine. This has already been realised in other areas of pathology, such as oncology [
54], where multi-modal foundation AI models have been developed. Multi-modal AI such as vision-language models, which integrate images (e.g., DP slides) and textual data, have been applied in other image-language domains like radiology [
55], and have been applied to MASH to identify steatosis and advanced fibrosis [
56].
At this time, major professional societies need to be cognizant of new developments in AI. We eagerly await guidance on standards for use of AI in diagnosis and reporting, as well as training of physicians in use of AI. Healthcare regulators such as the FDA are increasingly recognising the relevance of AI and DP, and we also look forward to increased adoption of these technologies in clinical trials and beyond.
FOOTNOTES
-
Authors’ contributions
DYZL and GBBG conceived of the study. All authors (DYZL, WQL, DSL, WKC, VWSW, GBBG) drafted the manuscript and approved of the final version.
-
Conflicts of Interest
The authors have no conflicts to disclose.
Figure 1.Envisioned digital pathology and artificial intelligence (AI) workflow. Image created with assistance of Google Gemini.
Table 1.Limitations to AI deployments in MASH digital pathology and their potential solutions
Table 1.
|
Limitations |
|
Potential solutions |
|
Technology |
Clinical relevance of more granular MASH and fibrosis parameters is not well established |
Further clinical validation |
|
AI methods are best studied in MASH, not well validated in mixed etiologies or other etiologies of liver disease |
Further clinical validation |
|
AI methods cannot compensate for inadequate sampling |
Emphasis on adequate sampling from biopsy operator |
|
Physician |
Pathologist trust in algorithm and output |
Utilizing AI as an assistant rather than as a replacement for a pathologist |
|
Performing local validation of AI models |
|
Healthcare system |
Requires additional workflows (e.g., slide digitalization, additional processing step) |
Training of laboratory staff, centralization of digital pathology and AI services |
|
Requires investment in new equipment |
Abbreviations
artificial intelligence-based measurement of non-alcoholic steatohepatitis histology
European Medicines Agency
Food and Drug Administration
hepatic venous pressure gradient
metabolic dysfunction associated steatohepatitis
metabolic dysfunction associated steatotic liver disease
non-alcoholic fatty liver disease
non-alcoholic steatohepatitis clinical research network
second harmonic generation
two photon excitation fluorescence
REFERENCES
- 1. Danpanichkul P, Suparan K, Diaz LA, Fallon MB, Chen VL, Namsathimaphorn K, et al. The rising global burden of MASLD and other metabolic diseases (2000-2021). United European Gastroenterol J 2025;13:1141-1154.
- 2. Younossi ZM, Kalligeros M, Henry L. Epidemiology of metabolic dysfunction-associated steatotic liver disease. Clin Mol Hepatol 2025;31:S32-s50.
- 3. Sanyal AJ, Jha P, Kleiner DE. Digital pathology for nonalcoholic steatohepatitis assessment. Nat Rev Gastroenterol Hepatol 2024;21:57-69.
- 4. Ratziu V, Hompesch M, Petitjean M, Serdjebi C, Iyer JS, Parwani AV, et al. Artificial intelligence-assisted digital pathology for non-alcoholic steatohepatitis: current status and future directions. J Hepatol 2024;80:335-351.
- 5. Kleiner DE, Brunt EM, Van Natta M, Behling C, Contos MJ, Cummings OW, et al. Design and validation of a histological scoring system for nonalcoholic fatty liver disease. Hepatology 2005;41:1313-1321.
- 6. Bedossa P, Poitou C, Veyrie N, Bouillot JL, Basdevant A, Paradis V, et al. Histopathological algorithm and scoring system for evaluation of liver lesions in morbidly obese patients. Hepatology 2012;56:1751-1759.
- 7. Angulo P, Kleiner DE, Dam-Larsen S, Adams LA, Bjornsson ES, Charatcharoenwitthaya P, et al. Liver fibrosis, but no other histologic features, is associated with long-term outcomes of patients with nonalcoholic fatty liver disease. Gastroenterology 2015;149:389-397.e10.
- 8. Sanyal AJ, Van Natta ML, Clark J, Neuschwander-Tetri BA, Diehl A, Dasarathy S, et al. Prospective study of outcomes in adults with nonalcoholic fatty liver disease. N Engl J Med 2021;385:1559-1569.
- 9. Loomba R, Ratziu V, Harrison SA. Expert panel review to compare FDA and EMA guidance on drug development and endpoints in nonalcoholic steatohepatitis. Gastroenterology 2022;162:680-688.
- 10. Park CC, Nguyen P, Hernandez C, Bettencourt R, Ramirez K, Fortney L, et al. Magnetic resonance elastography vs transient elastography in detection of fibrosis and noninvasive measurement of steatosis in patients with biopsy-proven nonalcoholic fatty liver disease. Gastroenterology 2017;152:598-607.e2.
- 11. Bedossa P, Dargère D, Paradis V. Sampling variability of liver fibrosis in chronic hepatitis C. Hepatology 2003;38:1449-1457.
- 12. Hui SL, Zhou XH. Evaluation of diagnostic tests without gold standards. Stat Methods Med Res 1998;7:354-370.
- 13. Merriman RB, Ferrell LD, Patti MG, Weston SR, Pabst MS, Aouizerat BE, et al. Correlation of paired liver biopsies in morbidly obese patients with suspected nonalcoholic fatty liver disease. Hepatology 2006;44:874-880.
- 14. Juluri R, Vuppalanchi R, Olson J, Unalp A, Van Natta ML, Cummings OW, et al. Generalizability of the nonalcoholic steatohepatitis Clinical Research Network histologic scoring system for nonalcoholic fatty liver disease. J Clin Gastroenterol 2011;45:55-58.
- 15. Lackner C, Gouw ASH, Avancini Ferreira Alves V, Arola J, Bedossa P, Behling C, et al. Consensus position statements for the standardized application of histological grading and staging systems in MASH clinical trials. J Hepatol 2026;84:693-701.
- 16. Davison BA, Harrison SA, Cotter G, Alkhouri N, Sanyal A, Edwards C, et al. Suboptimal reliability of liver biopsy evaluation has implications for randomized clinical trials. J Hepatol 2020;73:1322-1332.
- 17. Wong GL, Yuen PC, Ma AJ, Chan AW, Leung HH, Wong VW. Artificial intelligence in prediction of non-alcoholic fatty liver disease and fibrosis. J Gastroenterol Hepatol 2021;36:543-550.
- 18. Forlano R, Mullish BH, Giannakeas N, Maurice JB, Angkathunyakul N, Lloyd J, et al. High-throughput, machine learning- based quantification of steatosis, inflammation, ballooning, and fibrosis in biopsies from patients with nonalcoholic fatty liver disease. Clin Gastroenterol Hepatol 2020;18:2081-2090.e9.
- 19. Taylor-Weiner A, Pokkalla H, Han L, Jia C, Huss R, Chung C, et al. A machine learning approach enables quantitative measurement of liver histology and disease monitoring in NASH. Hepatology 2021;74:133-147.
- 20. Sun W, Chang S, Tai DC, Tan N, Xiao G, Tang H, et al. Nonlinear optical microscopy: use of second harmonic generation and two-photon microscopy for automated quantitative liver fibrosis studies. J Biomed Opt 2008;13:064010.
- 21. Farzi M, McGenity C, Cratchley A, Leplat L, Bankhead P, Wright A, et al. Liver-Quant: feature-based image analysis toolkit for automatic quantification of metabolic dysfunction-associated steatotic liver disease. Comput Biol Med 2025;190:110049.
- 22. Heinemann F, Gross P, Zeveleva S, Qian HS, Hill J, Höfer A, et al. Deep learning-based quantification of NAFLD/NASH progression in human liver biopsies. Sci Rep 2022;12:19236.
- 23. Akbary K, Noureddin M, Yayun R, Tai D, Boudes P. Development of AI based fibrosis detection algorithm by SHG/TPEF microscopy for fully quantified liver fibrosis assessment in MASH. Liver Int 2025;45:e70258.
- 24. Iyer JS, Juyal D, Le Q, Shanis Z, Pokkalla H, Pouryahya M, et al. AI-based automation of enrollment criteria and endpoint assessment in clinical trials in liver diseases. Nat Med 2024;30:2914-2923.
- 25. PharmaNest provides high-resolution, single-fiber phenotypic quantification of fibrosis from digital images of conventional stained slides for clinical and pre-clinical studies. Pharmanest web site, <https://www.pharmanest.com/technology>. Accessed 28 Feb 2026.
- 26. Liu F, Goh GB, Tiniakos D, Wee A, Leow WQ, Zhao JM, et al. qFIBS: an automated technique for quantitative evaluation of fibrosis, inflammation, ballooning, and steatosis in patients with nonalcoholic steatohepatitis. Hepatology 2020;71:1953-1966.
- 27. Xu S, Wang Y, Tai DCS, Wang S, Cheng CL, Peng Q, et al. qFibrosis: a fully-quantitative innovative method incorporating histological features to facilitate accurate fibrosis scoring in animal model and chronic hepatitis B patients. J Hepatol 2014;61:260-269.
- 28. Wang Y, Vincent R, Yang J, Asgharpour A, Liang X, Idowu MO, et al. Dual-photon microscopy-based quantitation of fibrosis-related parameters (q-FP) to model disease progression in steatohepatitis. Hepatology 2017;65:1891-1903.
- 29. Leow WQ, Bedossa P, Liu F, Wei L, Lim KH, Wan WK, et al. An improved qFibrosis algorithm for precise screening and enrollment into non-alcoholic steatohepatitis (NASH) clinical trials. Diagnostics (Basel) 2020;10:643.
- 30. Liu F, Wei L, Leow WQ, Liu SH, Ren YY, Wang XX, et al. Developing a new qFIBS model assessing histological features in pediatric patients with non-alcoholic steatohepatitis. Front Med (Lausanne) 2022;9:925357.
- 31. Ren Y, Chang JPE, Tai D, Filozof CM, Kleiner DE, Sanyal AJ. Repeatability and reproducibility of artificial imaging-based digital pathology for the evaluation of liver fibrosis. Clin Gastroenterol Hepatol 2026;24:713-722.
- 32. Soon GST, Liu F, Leow WQ, Wee A, Wei L, Sanyal AJ. Artificial intelligence improves pathologist agreement for fibrosis scores in nonalcoholic steatohepatitis patients. Clin Gastroenterol Hepatol 2023;21:1940-1942.e3.
- 33. Abdurrachim D, Lek S, Ong CZL, Wong CK, Zhou Y, Wee A, et al. Utility of AI digital pathology as an aid for pathologists scoring fibrosis in MASH. J Hepatol 2025;82:898-908.
- 34. Pulaski H, Harrison SA, Mehta SS, Sanyal AJ, Vitali MC, Manigat LC, et al. Clinical validation of an AI-based pathology tool for scoring of metabolic dysfunction-associated steatohepatitis. Nat Med 2025;31:315-322.
- 35. Brunt EM, Clouston AD, Goodman Z, Guy C, Kleiner DE, Lackner C, et al. Complexity of ballooned hepatocyte feature recognition: Defining a training atlas for artificial intelligence-based imaging in NAFLD. J Hepatol 2022;76:1030-1041.
- 36. DDT-BMQ-000105, Artificial Intelligence-Based Histologic Measurement of NASH (AIM-NASH) System. U. S. Food & Drug web site, <https://force-dsc.my.site.com/ddt/s/ddt-project?ddtprojectid=127>. Accessed 8 Feb 2026.
- 37. Qualification opinion for Artificial Intelligence-Based Measurement of Non-alcoholic Steatohepatitis Histology in Liver Biopsies to Determine Disease Activity in NASH/MASH Clinical Trials. European Medicines Agency web site, <https://www.ema.europa.eu/en/documents/other/qualification-opinion-artificial-intelligence-based-measurement-non-alcoholic-steatohepatitis-histology-liver-biopsies-determine-disease-activity-nash-mash-clinical-trials_en.pdf>. Accessed 8 Feb 2026.
- 38. Watson A, Petitjean L, Petitjean M, Pavlides M. Liver fibrosis phenotyping and severity scoring by quantitative image analysis of biopsy slides. Liver Int 2024;44:399-410.
- 39. Fujiwara N, Matsushita Y, Tempaku M, Tachi Y, Kimura G, Izuoka K, et al. AI-based phenotyping of hepatic fiber morphology to inform molecular alterations in metabolic dysfunction-associated steatotic liver disease. Hepatology 2026;83:575-590.
- 40. Harrison SA, Alkhouri N, Ortiz-Lasanta G, Rudraraju M, Tai D, Wack K, et al. A phase IIb randomised-controlled trial of the FFAR1/FFAR4 agonist icosabutate in MASH. J Hepatol 2025;83:293-303.
- 41. Harrison SA, Bedossa P, Guy CD, Schattenberg JM, Loomba R, Taub R, et al. A phase 3, randomized, controlled trial of resmetirom in NASH with liver fibrosis. N Engl J Med 2024;390:497-509.
- 42. Harrison S, Taub R, Labriola D, Ren Y, Chng E, Tai D. Artificial intelligence to measure fibrosis change on liver biopsy in MAESTRO-NASH a phase 3 52-week serial liver biopsy study in 966 patients with NASH treated with resmetirom or placebo. J Hepatol 2023;78:S112.
- 43. Sanyal AJ, Lopez P, Lawitz EJ, Lucas KJ, Loeffler J, Kim W, et al. Tropifexor for nonalcoholic steatohepatitis: an adaptive, randomized, placebo-controlled phase 2a/b trial. Nat Med 2023;29:392-400.
- 44. Naoumov NV, Brees D, Loeffler J, Chng E, Ren Y, Lopez P, et al. Digital pathology with artificial intelligence analyses provides greater insights into treatment-induced fibrosis regression in NASH. J Hepatol 2022;77:1399-1409.
- 45. Yuan HY, Tong XF, Ren YY, Li YY, Wang XL, Chen LL, et al. AI-based digital pathology provides newer insights into lifestyle intervention-induced fibrosis regression in MASLD: an exploratory study. Liver Int 2024;44:2572-2582.
- 46. Bosch J, Chung C, Carrasco-Zevallos OM, Harrison SA, Abdelmalek MF, Shiffman ML, et al. A machine learning approach to liver histological evaluation predicts clinically significant portal hypertension in NASH cirrhosis. Hepatology 2021;74:3146-3160.
- 47. Kendall TJ, Chng E, Ren Y, Tai D, Ho G, Fallowfield JA. Outcome prediction in metabolic dysfunction-associated steatotic liver disease using stain-free digital pathological assessment. Liver Int 2024;44:2511-2516.
- 48. Field DT, Ren Y, Akbary K, Chng E, Tai D, Naoumov NV, et al. Effect of liver biopsy size on MASLD fibrosis assessment by second-harmonic generation/two-photon excitation fluorescence microscopy. JHEP Rep 2025;7:101449.
- 49. Zhang M, Chen S, Wu X, Zhou J, Wang T, Liu H, et al. Persistent steatosis correlates with decreased fibrosis regression during anti-HBV treatment in patients with chronic HBV infection. J Med Virol 2023;95:e29156.
- 50. Abdurrachim D, Wee A, Soon G, Idowu M, Kendall TJ, Lek S, et al. AI-assisted fibrosis scoring in MASH: exploring pathologist decision-making with an SHG-based AI digital pathology tool. JHEP Rep 2026;8:101752.
- 51. Lujan G, Quigley JC, Hartman D, Parwani A, Roehmholdt B, Meter BV, et al. Dissecting the business case for adoption and implementation of digital pathology: a white paper from the digital pathology association. J Pathol Inform 2021;12:17.
- 52. Hanna MG, Ardon O, Reuter VE, Sirintrapun SJ, England C, Klimstra DS, et al. Integrating digital pathology into clinical practice. Mod Pathol 2022;35:152-164.
- 53. Digital Pathology: The Top Four Cost Saving Use Cases in Pathology Labs. PathAI web site, <https://www.pathai.com/blog/digital-pathology-the-top-four-cost-saving-use-cases-inpathology-labs>. Accessed 14 May 2026.
- 54. Tripathi A, Waqas A, Schabath MB, Yilmaz Y, Rasool G. HONeYBEE: enabling scalable multimodal AI in oncology through foundation model-driven embeddings. NPJ Digit Med 2025;8:622.
- 55. Dutta N, Bose K, Syailendra E, Chu L, Gupta P. Vision-language models in diagnostic imaging: review of technical advances, clinical validation, and practical deployment. Int J Med Inform 2026;208:106227.
- 56. Zhang Y, Liu H, Sheng B, Tham YC, Ji H. Preliminary fatty liver disease grading using general-purpose online large language models: ChatGPT-4 or Bard? J Hepatol 2024;80:e279-e281.