Automated liver histology scoring brings quantitative precision to MASH and MASLD
The burden of MASLD and MASH has outgrown the limits of manual histology. In liver biopsies, the central problem is no longer whether steatosis, inflammation, ballooning, and fibrosis matter clinically; it is whether those features can be measured with enough objectivity and consistency to support diagnosis, risk stratification, and treatment monitoring. A fully automated deep-learning pipeline built on nnU-Net now attempts to do exactly that, converting routine H&E slides into slide-wide quantitative readouts.
The system, called qHisto, was trained and applied to H&E sections from 45 mice and 44 humans with MASH/MASLD. Its design is intentionally broad: it segments liver tissue and several histological components relevant to the MASH score, then translates those segmentations into numerical outputs. The result is not simply another image-analysis tool for a single lesion class. It is a structured attempt to mirror established semiquantitative pathology while also adding measurements that manual review cannot reliably generate, such as fat droplet size distributions, nuclear aspect ratio, and spatial heatmaps.
One sentence captures the clinical pressure behind this work: conventional scoring remains indispensable, yet it is still coarse, subjective, and slow.
Whole-slide segmentation tracked fat, fibrosis, inflammation, and ballooning
qHisto was applied across different tissue models to segment features relevant to liver disease severity, including fat, fibrosis, inflammation, nuclei, sinusoids, vessel structures, ballooning, microvesicular steatosis, and glycogenated nuclei. The pipeline’s segmentation performance was supported by Dice and F1 scores across the trained models, indicating that the algorithm could recognize and delineate the structures needed for downstream quantification.
The practical advantage of this approach is spatial resolution. Conventional semiquantitative histology compresses a great deal of tissue heterogeneity into ordinal categories, often masking how lesions are distributed across a slide. By contrast, qHisto provides pixel-level quantification. For a disease defined by patchy fat accumulation, focal inflammation, and uneven fibrotic remodeling, the capacity to measure where abnormalities sit within the tissue is not cosmetic; it is biologically relevant.
That matters because MASH does not injure the liver in a uniform way. Fat droplets enlarge some hepatocytes more than others, inflammatory foci cluster unevenly, and fibrosis evolves along architectural boundaries rather than appearing as a simple global average.
Quantitative outputs tracked closely with standard histology and MRI fat fraction
Validation anchored the new pipeline against conventional histology and imaging. For steatosis, qHisto-derived fat area closely tracked both human and murine reference measures. In the pooled human and mouse cohorts, fat area showed a strong positive correlation with hepatic fat fraction on MRI, with R = 0.87 and p = 0.74 × 10−17. When the data were evaluated against conventional histological steatosis grading, the correlation was even stronger, reaching R = 0.91 with p = 0.91 × 10−28 in the combined dataset.
In the murine cohort, the temporal pattern of qHisto-derived fat area also followed the MRI trajectory more faithfully than conventional scoring. Fat values were low at baseline, rose through the middle time points, and then declined later, while the manual score plateaued earlier. That divergence is clinically important because a score that flattens too soon can miss biologically meaningful change. A quantitative pipeline that continues to move with tissue burden may offer a more sensitive tool for longitudinal studies and therapeutic readouts.
Fibrosis measurements also aligned with established pathology, although the signal was more modest than the steatosis correlations. Quantified fibrosis area correlated with conventional fibrosis scoring at R = 0.51 and p = 0.47 × 10−5. Separation between score groups improved for clinically meaningful contrasts, including 0 versus 1–2 and 0–1 versus 2, with AUC values of 0.80 and 0.83. In the murine cohort, the fibrosis correlation remained positive, while the human subset showed a weak, non-significant association. That split suggests the pipeline is capturing fibrosis biology, but also that performance may depend on cohort composition, sampling, or the limited number of human cases available for specific comparisons.
Nuclear deformation and droplet geometry added a new layer of tissue biology
The most interesting part of qHisto is not its ability to reproduce what expert readers already know. It is the ability to quantify what readers usually infer but do not measure directly. Nuclear aspect ratio showed a strong positive correlation with fat area across the pooled human and murine data, with R = 0.80 and p = 3.32 × 10−20. The proportion of elongated nuclei also correlated strongly with fat burden, reaching R = 0.78 and p = 4.28 × 10−18.
Biologically, this is plausible. Large fat droplets distort cellular architecture, displacing cytoplasmic contents and deforming nuclei as hepatocytes enlarge. The data suggest that nuclear deformation is driven more by the presence of larger droplets than by total fat amount alone. In other words, tissue mechanics may depend on droplet morphology, not just fat volume. That kind of insight is difficult to capture with ordinal scoring, which counts severity but not shape.
The droplet-size analysis adds another dimension. The skewness of fat droplet diameter separated murine time points into distinct groups, with increased skewness appearing later in the disease course. Severe steatosis cases showed a visibly different droplet-size distribution from milder cases. Human cases could not be analyzed in the same longitudinal way because only one time point was available per patient, limiting the ability to trace temporal evolution in the clinical cohort.
Quantitative histology may support trial endpoints, but the human dataset remains limited
qHisto was also used to derive components of the broader MASH score, including microvesicular steatosis, fibrosis, lobular inflammation, portal inflammation, ballooning, and glycogenated nuclei. Inflammation analysis improved when the minimum number of cells per focus was set to 5 rather than 3, with a correlation of R = 0.70 and p = 0.77 × 10−11. Ballooning correlation with conventional scoring was more modest at R = 0.47, while the fat component remained the strongest and most reproducible feature in the pipeline.
The implications are straightforward for pathology workflows and clinical research. Automated quantification can reduce inter-reader variability, speed up repetitive slide review, and generate continuous data where manual systems currently assign bins. For drug development in MASH, where histology often anchors eligibility and response assessment, a reproducible slide-wide method could strengthen signal detection and standardize central reading.
The limitations are equally direct. The human cohort was relatively small, some features performed better than others, and several novel descriptors were explored in mice without matching longitudinal human data. The fibrosis signal was weaker in humans than in mice, and microvesicular steatosis did not discriminate well between lower conventional scores. qHisto therefore looks most mature as a quantitative complement to expert pathology rather than a replacement for it. Its value lies in turning a subjective tissue review into a reproducible measurement framework, while leaving open the question of how these richer metrics will perform in larger multicenter cohorts and in therapy-response studies.