This study utilized tissue samples from 145 piglets reared in a conventional indoor intensive production herd in Denmark. Included animals were not subjected to litter equalization or cross-fostering, and all piglets were weaned at 26 days of age. The animals were randomly selected from a cohort of 2,500 piglets representing individuals both with and without signs of gastrointestinal disease. The PWD disease prevalence in the study was 43%. The piglets were stratified into five different age groups: 4, 14, 25, 49, and 67 days.

Between July 2023 and March 2024, samples were collected in 11 batches, each containing 10–15 animals. On the morning of sampling, all piglets were transported for approximately 1.5 h from the farm to Frederiksberg Campus, University of Copenhagen. The piglets were anesthetized via intramuscular injection of Zoletil (0.1 mL/kg) and euthanized by intracardiac injection of an overdose of pentobarbital. Following evisceration, full-thickness samples of mid-jejunum were collected from all animals, while samples from duodenum and ileum were taken from a subset of 75 animals representing all age groups. Duodenum was identified as the segment between the pylorus of the stomach and the oral aspect of the duodenocolic fold. Jejunal samples were taken from mid-jejunum, approximately midway between the aboral edge of the duodenocolic fold and the ileocecal fold. Ileal samples were identified as the segment between the oral edge of the ileocecal fold and the ileal inlet to cecum. The samples were opened, gently rinsed in isotonic saline, and pinned onto styrofoam boards before being immersed in 10% neutral buffered formalin. All samples were secured in fixative within 15 min of euthanasia.

Samples were immersion-fixed in 10% neutral buffered formalin for 3 days. After fixation, tissues were trimmed and processed through graded concentrations of ethanol and xylene, embedded in paraffin, and samples of 5 μm were stained with hematoxylin and eosin (H&E) using standard protocols as previously described (6).

Stained histological slides were digitized for image analysis by whole-slide scanning with a Zeiss Axioscan Z1 scanner, using a 20x/0.8 Plan-Apochromat objective. Software settings were adjusted specifically for brightfield microscopy of 5 μm H&E stained intestinal tissues.

Image pre-processing and ML model training were conducted on a Lenovo ThinkPad laptop with an Intel Core i5 processor, to mimic the commonly available hardware setup in veterinary research settings. Model-training and image analysis was conducted using the open-source image analysis software QuPath [version 0.5.1 (21),]. Model training was performed exclusively using jejunum samples to ensure consistency in tissue structure and morphology. Prior to analysis, all tissue sections were individually color-normalized to H&E using the built-in “Estimate Stain Vectors” tool. To achieve complete segmentation of the intestinal wall layers, a 3-step separation model was developed (Figure 1). All settings adjusted for model development are mentioned in the following paragraphs. If not mentioned, the settings were left as default.

In the first step, tissue was automatically identified and annotated using a pixel-based thresholding approach. This was achieved by analyzing the average RGB values (i.e., Red-Green-Blue values) of each pixel at a resolution of 7.03 μm/px, with a Gaussian prefilter and a smoothing sigma of 2.0 applied. A threshold value of 205 was used as cutoff, with pixels below this value classified as tissue and those above classified as background (Figure 1B). In order to measure the tissue area, the classification was converted to an object using the “create objects” function, with minimum object size of 10,000 μm² and a minimum hole size of 100,000 μm² to avoid misclassification of vascular structures and edematous areas.

In the second step, a machine learning-based pixel classifier, utilizing the default random trees algorithm, was trained to differentiate the intestinal mucosa from the underlying submucosa and muscular layers (Figure 1C). Classification was performed based on the RGB and Hematoxylin channels of the image, with a resolution of 7.03 μm/px. The classifier employed features of Gaussian, Laplacian of Gaussian, Weighted Deviation, and Hessian Determinant, with scaling factors of 1.0 and 2.0, a selection based on published guidelines (22). Training images were blinded with regards to disease status and added ad hoc by continuous inclusion of problematic slides, and the final model was developed using 236 manual annotations of “mucosa” and “submucosa/muscularis” from 15 distinct animals (n₄ = 2; n₁₄ = 2; n₂₅ = 2; n₄₉ = 8; n₆₇ = 1). All annotations were made by the main author (CBB) using the built-in brush-tool in QuPath to provide the software with examples of mucosal and submucosal tissues, respectively. Area extraction was achieved by applying the same approach as described for Step 1.

In the final step, another pixel classifier utilizing a random trees algorithm was trained to differentiate the epithelium from the lamina propria within the mucosal area identified previously by the model (Figure 1D). As for the second step, classification was based on the RGB and Hematoxylin channels, but at a higher resolution of 3.51 μm/px. To address the increased complexity of the structures, the classifier incorporated additional features including Gaussian, Laplacian of Gaussian, Weighted Deviation, Gradient Magnitude, Structure Tensor Coherence, and Hessian Determinant, all with a scaling factor of 1.0. The training set comprised eight distinct tissue samples included ad hoc, and 172 manual annotations (n₄ = 0; n₁₄ = 1; n₂₅ = 2; n₄₉ = 3; n₆₇ = 2) made in a similar fashion as for Step 2. For area extraction a minimum object size and minimal hole size of 1,000 μm² was applied.

To account for potential age-related differences in tissue morphology, Steps 2 and 3 were repeated, and separate classifiers were trained on new images using age-balanced datasets with all age-groups represented in the training data. The age-balanced training for Step 2 again included 15 tissue samples and 236 annotations (n₄ = 2; n₁₄ = 2; n₂₅ = 2; n₄₉ = 8; n₆₇ = 1), while Step 3 utilized 10 tissue samples and 285 annotations (n₄ = 2; n₁₄ = 2; n₂₅ = 2; n₄₉ = 3; n₆₇ = 2). The classifier specifications, including feature sets and scaling factors, remained unchanged.

The performance of the trained models, including both the Ad Hoc trained and the Age-balanced models, was assessed by comparing the areas predicted by each step of the classification process (A_model) to manually annotated areas (A_manual, applied as ground truth), across all 145 jejunal samples. In general terms, the above mentioned samples (15, 8, and 10 samples respectively) were selected as training sets, while the whole dataset of 145 animals were used for validation. No separate dataset for a final testing phase was available in this study. For each sample, a 5 mm section containing longitudinally sectioned crypts of Lieberkühn was selected for validation. The length of the section was determined by measuring along the lamina muscularis mucosa using the polyline tool. The relative deviation from the manual annotations was calculated as follows:

Differences in relative deviation across age groups were visualized using a modified Bland–Altman plot and categorized according to level of agreement. Agreement levels were defined as follows: RD < 5% = Very good, RD = 5–10% = Good, RD = 10–20% = Fair, and RD > 20% = Poor. The performance of each classification step was further evaluated within the five age groups by calculating the Pearson correlation coefficient between manually annotated areas and areas provided by the trained models (23).

To assess spatial agreement between manually annotated regions and AI-generated segmentations, Intersection over Union (IoU) was calculated. This metric quantifies the proportion of overlap between the two areas relative to their combined area, thereby verifying that segmented regions were co-located rather than merely adjacent. IoU values were interpreted using commonly applied thresholds, with values >0.50 considered acceptable, >0.75 considered good, and >0.90 considered near-perfect. The QuPath script used for IoU calculation is provided in the Supplementary material. In addition to the quantitative evaluation, a visual inspection of classification errors was performed to identify recurring misclassifications and limitations in the applied methodology. Tissue structures that were consistently misclassified in the different steps of the classification process were recorded in order to guide potential adjustments in future training iterations.

To assess the translatability and adaptability of the Ad Hoc trained model to other intestinal regions than the jejunum assessed above, three samples from both duodenum and ileum from each of the five age groups were evaluated. A section measuring 2.5 mm of each sample was evaluated as described above. The selection of segments for analysis followed the same principles as for jejunal samples, but actively avoiding segments containing Peyer’s patches in ileum. The model’s performance on these samples was compared to its performance on the jejunum samples from the same piglets, providing insight into its ability to generalize across different anatomical regions in the small intestine.

All statistical analyses were performed using R Statistical Software [R version 4.4.0 (2024-04-24)] (24). Prior to data analysis all parameters were evaluated for normality by visual inspection of histograms.

The study was approved by the Danish Animal Inspectorate (License no: 2022-15-0201-01324).

Step 1 (tissue detection) demonstrated high accuracy and consistency across all age groups. Mean relative deviations ranged from 2.9 to 10.7%, with the highest deviations observed in the youngest age group (4 days), and gradually increasing performance with increasing age (Figure 2). The Pearson correlation coefficient for tissue detection was very high (0.97), indicating a highly correlated relationship between AI-generated and manually annotated areas across all age groups (Table 1; Figure 3A). When categorized by agreement levels, Step 1 performed robustly, achieving 80% “very good” agreement in the oldest group (67 days) and 73% in 49-day-old piglets. However, in the youngest age group (4 days), only 30% of cases achieved “very good” agreement, and 30% were classified as either “fair” or “poor” (Table 2). These results show that tissue detection was consistent overall, but with a reduced accuracy for the younger age groups. When assessing the IoU values, the areas provided by the model showed consistently high colocalization with manual annotations (overall μ = 0.91, sd = 0.09). Mean IoU values were above 0.90 for all age groups, except the 4-day olds, thus exceeding the threshold generally interpreted as “near-perfect” alignment (Table 1).

Mucosal segmentation exhibited greater variability compared to Step 1. Relative deviations varied across age groups, with mean deviations as high as 24.6% in the youngest piglets but improving to 7.6–10.0% in older age groups (Table 1; Figure 2). The Pearson correlation coefficients dropped similarly, with an overall coefficient of 0.67 corresponding to strong linearity between AI-generated and manual annotations (Figure 3B, Supplementary Figure 1). Categorization of agreement level revealed a notable age-dependent trend in the performance of Step 2. For example, in the youngest group (4 days), only 10% of cases achieved “very good” agreement, while 40% were categorized as “poor.” Performance improved significantly in older groups, reaching 57% “very good” and only 7% “poor” at 67 days (Table 2). The spatial alignment of annotated areas mirrored the trend for the other performance parameters, yielding an overall IoU of 0.78 ± 0.19, and a positive association between IoU and increasing age (Table 1).

The final step in the segmentation model presented similar challenges as those in Step 2 with an overall mean deviation of 12.6%, and a large variation was seen between the different age groups ranging from 5.6–30.1%. As with the previous steps, the largest variation was seen in the youngest age group (4 days), which also affected the Pearson correlation coefficient for this age to drop to 0.49 (moderate linear correlation). The Pearson correlation coefficients for the remaining age groups were considerably stronger, varying from 0.77–0.96 (Table 1, Figure 3C; Supplementary Figure 2). When evaluating the categorized data, inferior performance was again observed in the youngest age groups, especially in the 4-day-old piglets, where more than half of the samples (55%) were assigned “poor.” The trend was almost opposite in the oldest age group, where 47% were categorized as “very good” agreement, and only 1 piglet ended up in the “poor” agreement category (Table 2). Despite the deviation in area size, the alignment between manually annotated and model-generated areas remained high, with an overall IoU of 0.92 ± 0.11.

When comparing the performance of the Ad Hoc trained model to the Age-balanced model, notable differences emerged in Step 3. For Step 2 (mucosal segmentation), the overall performance was comparable between the two models (Table 1; Figures 2B, 3B). However, in Step 3 (epithelial segmentation), the Ad Hoc model consistently outperformed the Age-balanced model across all age groups. For example, in 49-day-old piglets, the Ad Hoc trained model achieved 59% “very good” agreement, while the Age-balanced model only achieved “very good” agreement in 7% of cases, and 70% was categorized as “poor” agreement (Table 2). Relative deviation, Pearson correlation coefficient, and IoU values mirrored these results, with the Age-balanced model exhibiting consistently larger deviations than the Ad Hoc model (Table 1; Figure 3C). The clear age-dependent decrease in relative deviation seen for the final step of the Ad Hoc trained model was not evident for the age-balanced model (Figure 2C).

The performance of the Ad Hoc model was evaluated on duodenum and ileum samples to assess its translatability across different segments of the small intestine. Tissue identification (Step 1) demonstrated consistent performance across all three anatomical regions (Supplementary Figure 3A). However, while mucosal segmentation (Step 2) showed poor results in the youngest age group for both duodenum and ileum, the performance was generally good in older age groups, with only slightly inferior performance in duodenum and ileum compared to the associated jejunum samples (Supplementary Figure 3B). Epithelial segmentation (Step 3) exhibited stable accuracy across all age groups in duodenum, comparable to jejunal performance. In contrast, ileum displayed an age-dependent improvement in accuracy, with better results in older piglets. Post-weaning, the relative deviation averaged −5% for duodenum, −4% for jejunum, and 0% for ileum, reflecting “very good” performance for Step 3 across all segments in older animals (Supplementary Figure 3C).

During the comparison of AI-generated and manually annotated areas, a qualitative assessment of classifier errors was performed. Common segmentation errors were classification step-specific, although some mistakes were shared between Step 2 and 3 of the classification process.

In Step 2, the segmentation of the tissue into mucosa and submucosa relied on the training of a random trees algorithm. A predominant error was observed in younger piglets (4–21 days), where vacuolated fetal-type enterocytes in the epithelial lining were frequently misclassified as part of the submucosal tissues due to their distinct texture and fainter staining intensity (Figure 4B).