Metabolic Dysfunction-Associated Steatotic Liver Disease (MASLD) is the most common cause of chronic liver disease worldwide, with a global prevalence of 25% in the adult population, 60% in patients with type 2 diabetes and 80% in obese individuals (1, 2). Only a small proportion of MASLD cases progress to cirrhosis and severe complications, but biomarkers predicting patients at risk are still lacking. Conversely, in patients where steatosis is complicated by inflammation and cellular damage (i.e., MASH - Metabolic Dysfunction-Associate Steatohepatitis), the possibility of severe liver diseases increases proportionally with the amount of fibrosis, as estimated by the pathological stage (3, 4). In keeping with this, the histopathological reduction of fibrosis is one of the main endpoints for clinical trials on MASH patients (5–7). However, the histopathological assessment of fibrosis stage in chronic liver disease, including MASH (8, 9), is a descriptive and semi-quantitative process with moderate reproducibility even among expert liver pathologists (10, 11). Moreover, the limited categories used for pathological staging scarcely illustrate a disease driven by a continuum of injury, resulting in limited sensitivity to reflect changes in disease severity over time (11).

Approaches based on whole slide digital images (WSI) and Artificial Intelligence (AI) have been applied successfully to the evaluation of morphological features of MASH, including fibrosis (12–18). Notably, two methods have shown clinically meaningful results. Collagen Proportionate Area (CPA), a quantitative assessment of collagen area expressed as the ratio of collagen-stained pixels over full-biopsy-pixels, demonstrated its superiority over semi-quantitative staging in predicting clinical decompensation (19, 20). Second Harmonic Generation (SHG) analysis, which focuses on the possibility to visualize collagen allowing its quantification on a continuous scale and evaluation of spatial relationship with the surrounding cells (14, 21), demonstrated its superiority over pathological staging in detecting treatment-associated reduction of fibrosis (15). These new methods, allowing to characterize minimal difference in the quantitative features of the fibrous tissue, pave the way to precision medicine also in hepatology (11).

The aim of this study is to test the clinical impact of an AI-based approach in characterizing liver fibrosis features in a series of MASH patients.

The study cohort consisted of 52 patients (Table 1). Among them, 9 underwent a post-treatment biopsy. Therefore, the total number of cases was 61. The results of the pathologists’ agreement in staging these cases are shown in Figure 1. The agreement was fair for F1a, F1b and F1c (respectively κ: 0.19, 0.15 and 0.17), moderate for F2 (κ: 0.43) and F3 (κ: 0.58) and substantial for F4 (κ: 0.73). The overall agreement was only moderate (κ: 0.44 Figure 1A, rising to κ: 0.55 when grouping all F1 classes into a single category, Figure 1B). After the adjudication, the series was represented by 9 F1 (a,b,c: 4,1,4), 19 F2, 28 F3 and 5 F4. These diagnoses were then considered as the ground truth.

Histopathological characterization of fibrosis is the only parameter that correlates with the evolution of steatohepatitis (5–7). A reduction in fibrosis is one of the goals of the treatments for steatohepatitis, as well as one of the parameters used by the FDA for drug approval (25). Nonetheless, the evaluation of fibrosis modification under the microscope has some limitations. First, the assessment is a highly subjective process: numerous studies have demonstrated good but not exceptional agreement even among expert hepato-pathologists (26). Then, the system in use classifies into defined categories, similar to impermeable compartments, a process, the appearance, and modulation of fibrosis, which in nature appears as continuous (27). Finally, liver fibrosis remodeling is a slow event that might not be depicted properly by the few categories of current staging systems. To overcome these limitations, some studies started using AI to quantify fibrosis in a more objective way and as a continuous parameter, or to identify specific parameters that might not be defined by histopathology, for example the entropy of collagen fibers (13, 16, 28, 29).

The results of the present study, conducted on MASH biopsy confirm that AI-based evaluation characterizes at the deepest level, compared to classic histopathological evaluation, the fibrotic process occurring in patients with MASH. In particular, we showed that cases consistently labeled as belonging to the same histopathological stage might show significant differences in collagen amount. Thus, cases classified as F3 might present values identical to a F2- or to a F4- cases. This infinitesimal evaluation of the collagen amount likely represents the result of the multi-resolution approach we adopted in the study. Indeed, the AI we designed combined features related to texture and intensity obtained at two clearly separated spatial scales: single pixel and ROI (32x32pixel). Interestingly, Taylor-Weiner et al. (16), using a similar approach, a deep-learning method able to measure the amount and distribution of fibrosis across the whole biopsy at different resolutions, proved a significant heterogeneity across patients with the same stage. A significant heterogeneity across patients with the same stage was also proved by Wang et al. (29) and Forlano et al. (13) using Collagen Proportionate Area (CPA). Interestingly enough the values of CPA reported by these two studies for specific MASH stages (F2: 3.3 and 2.1%; F3: 6.8 and 5.5%; F4: 12.4 and 11.1%) overlap those of ECA in our study (F2: 2.6%; F3: 5.7% and F4: 10.9%).

The present study also confirmed that AI allows a more accurate identification of minimal changes conditioned by the therapy in MASH patients. In fact, while histopathology identified stage changes in 53% of cases, the congruous combination of two AI features representing the amount and entropy of collagen highlighted the presence of changes in 76% of the cases. These findings are strictly in keeping with the few similar data available in the literature. Naoumov et al. (15) proved a change in stage in 97% of cases using digital quantification on SHG as compared to 40% detected by histopathology. In specific, stage regression was observed in 69% of F2 and F3 cases treated with the optimal dose of drug. Taylor-Wayner et al. (16) calculated 27 and 40% of responders using, respectively, a rigorous or less-stringent thresholds as compared to none detected by histopathology. In our series, AI proved a stage regression in 53% of cases; histopatology in 29%; weight loss greater than 10%, a clinical feature that has been associated with improvements in liver fibrosis in MASH (24), in 12%. We also showed that the reduction in absolute terms of collagen was frequently accompanied by a reduction of entropy of collagen (EnC). This latter is a textural parameter that translates the randomness of Sirius Red optical density values with respect to its neighborhood in terms of the intensity distribution. Low entropy values correspond to more uniform and homogeneous image thus to more grouped collagen fibers. Interestingly enough, our findings overlap the description of regressing fibrosis reported by Sanyal et al. (7) characterized by the narrowing or even disappearance of the fibrotic bridge; as well as that described by Naoumov et al. (15) showing that after treatment septa became thin and compact, with sharp borders. Once inserted into this framework, our results, although limited to a small series of post treatment biopsy, do not appear random but rather representative of the important role played by AI in highlighting changes caused by medical therapy.

Finally, we showed that some post treatment liver biopsy totally unchanged under the microscope, have only minimal changes also in the parameters extracted by AI. This result mirrors those of a SHG study showing that fibrosis regression starts in a piecemeal manner (15). Interestingly enough, a recent paper, based on the spatial information returned by SHG proved, in a murine model of steatohepatitis, that this minute fibrosis regression is induced by a reduction in food amount (30).

In conclusion, our data confirm that AI allows a better characterization of the fibrosis process by its true, continuous, and non-categorical nature. Of clinical importance is also the possibility, guaranteed by AI algorithms, of identifying among the cases that the pathologist would have judged as unchanged, those in which there was actually a response to treatment. Finally, it does not appear irrelevant to observe that there is a certain percentage of cases where the objective parameters generated by AI highlight an initial process of modification of the fibrosis. Taken together, the intrinsic possibilities of AI, pave the way for renewed staging system, where the information generated by pathological description (site of fibrosis; formation of septa) would benefit of an integration with data generated by AI.