Machine-learning for quantitative histopathology of piglet intestinal tissues: Challenges with limited training data

Cecilie Brandt Becker^*Mette Sif HansenSøren Saxmose NielsenHenrik Jensen

• University of Copenhagen, Copenhagen, Denmark

Introduction: Use of Machine learning (ML) is rapidly expanding in histopathology, offering the potential to reduce interobserver variability and improve quantitative assessment. However, large datasets and computational resources commonly used in toxicology and human medicine are often unavailable to the veterinary pathologist. This study aimed to evaluate the feasibility and limitations of applying supervised ML on histopathological samples with limited training data, exemplified by training an ML model to segment the intestinal wall into its histological layers. Materials and Methods: The study included 145 piglets from five age groups (4, 14, 25, 49, and 67 days). Full-wall samples from duodenum, jejunum and ileum were collected post-mortem, stained with H&E and digitized. A three-step ML model was trained on 8-15 images: Step 1 identified tissue, Step 2 segmented mucosa from submucosal layers, and Step 3 separated lamina propria from epithelium. Model performance was assessed by comparing AI-generated areas to manual annotations, calculating relative deviation, categorized agreement levels, Intersection over Union, and Pearson correlation coefficients. Qualitative error analyses were used as directions for future training options. Results: A three-step separation model was successfully developed, but showed a significant amount of age-related performance variation, depicted as larger inaccuracies in samples from the younger age-groups, reflecting additional tissue heterogeneity from immature morphology. Classification errors could be categorized into intrinsic limitations (e.g., thresholding issues in tissue identification) and training deficits (e.g., misclassification of goblet cells and crypt abscesses), of which only the latter category could be corrected by adding additional training data. Conclusion: This study demonstrates the feasibility of ML-based histopathology with limited sample sizes, providing a viable option for veterinary pathologists. Models trained on small datasets require careful supervision, with special emphasis on age-diverse tissue heterogeneity and overfitting. In these cases, ML should be seen as a tool to augment, not replace, expert oversight, ensuring reliable and reproducible quantitative histopathological measures.