<p>Liver fibrosis, a major histopathological indicator of chronic liver injury, is also a key feature of metabolic dysfunction-associated steatohepatitis. Its quantitative assessment in preclinical toxicology is frequently inconsistent and subjective. This study aimed to develop and validate multi-scale, patch-based convolutional neural network classification algorithms for automated fibrosis quantification in a carbon tetrachloride (CCl₄)-induced mouse model. We sought to determine the optimal patch size for accurate predictions. Accordingly, male C57BL/6 mice (n = 19) were categorized into the following three groups: vehicle control (n = 5), high-fat diet (HFD) and CCl₄ positive control (n = 9), and HFD and CCl₄ with elafibranor (ELA) treatment (n = 5). Liver tissues were stained with Sirius-red, digitized as whole slide images, and cropped into patches of 32 × 32, 64 × 64, or 128 × 128 pixels. Each algorithm was trained, validated, and tested in an 8:1:1 ratio over 40 epochs with a batch size of 32 to classify fibrotic, normal, and background regions. All models performed robustly, with validation accuracies exceeding 98% and F1-scores above 0.96. Particularly, the 32 × 32 model exhibited the highest correlation with pathologist’s measurements (Spearman’s <i>r</i> = 0.9609; <i>p</i> &lt; 0.05) and the most accurate estimation of absolute fibrotic area compared to expert assessments. This model also accurately detected the antifibrotic effects of ELA. These findings establish that the 32 × 32 patch-based classification approach provides a rapid, reproducible, and objective method for liver fibrosis quantification in preclinical toxicology, with strong potential for integration into digital pathology workflows.</p> Graphical abstract <p></p>