ABSTRACT

Background and Aim: Conventional semiquantitative histopathological scoring of hepatic fibrosis is affected by interobserver variability and provides limited quantitative information on the spatial distribution and progression of collagen deposition. Machine learning (ML)-based image analysis may provide a reproducible approach for objective fibrosis assessment in veterinary pathology. This study compared three open-source ML models, Trainable Weka Segmentation (TWS), Ilastik, and Fiber-ML, for segmentation and quantitative assessment of bovine liver fibrosis and evaluated their diagnostic performance, agreement with METAVIR staging, and computational efficiency.

Materials and Methods: This cross-sectional observational study included liver samples from 35 cattle diagnosed with hepatic fibrosis according to the clinical sign records. Masson’s trichrome-stained sections were classified according to METAVIR stages F0–F4 (F0 represents normal liver samples, n=7). A total of 525 images were obtained from perisinusoidal, pericentral, and periportal regions. Images from 25 animals (375 images) constituted the training/validation dataset with five-fold cross-validation, whereas images from 10 animals (150 images) formed an independent testing dataset. A similar dataset was annotated by experts to generate manual ground-truth segmentation, and images from the same animal appeared in both training and testing datasets. Model performance was assessed using collagen proportionate area (CPA), Dice similarity coefficient (DSC), sensitivity, specificity, accuracy, correlation with METAVIR scores, and computational-resource requirements.

Results: Ilastik and Fiber-ML generally achieved greater segmentation agreement than TWS (p ≤ 0.05), with DSC values reaching 0.91 and 0.92, compared to 0.87 in TWS, depending on the hepatic region. Overall model pixel-level segmentation accuracy ranged from 84.7% to 91.8%, with the highest performance observed in the periportal region. CPA measurements generated by the models showed strong positive correlations with METAVIR stages (ρ = 0.97–0.99). Fiber-ML required substantially less computational resources, while Ilastik recorded fewer software errors than TWS, including shorter training time and lower CPU and RAM utilization. All computational benchmarking was conducted under identical hardware and software conditions.

Conclusion: Open-source ML models enabled objective segmentation and quantitative assessment of bovine hepatic fibrosis. Ilastik and Fiber-ML demonstrated strong agreement with expert-derived reference segmentation and fibrosis staging, while Fiber-ML offered additional computational efficiency. These findings support ML-assisted digital histopathology as a future complementary tool for quantifying and spatially characterizing hepatic fibrosis in cattle. Further validation is required before their reproducibility and diagnostic applicability can be established.

Keywords: bovine liver fibrosis, collagen proportionate area, digital histopathology, hepatic fibrosis, machine learning, METAVIR staging, segmentation, veterinary pathology.