This retrospective study was performed in accordance with the Salerno Kidney Transplant Cohort Study, reported elsewhere [24], and approved by the local Institutional Ethical Committee. The Salerno Kidney Transplant Cohort Study is an open-ended longitudinal study, based on follow-up of kidney transplant recipients self-referring to the Nephrology Unit of the Salerno University Hospital. The study included KTRs who underwent a biopsy of the lower gastrointestinal tract as per protocol or within a diagnostic path for gastrointestinal symptoms. Colon cancer screening protocols for kidney transplant patients followed the same guidelines for the general population. This included colonoscopy screening for patients over 50 years old, typically repeated every 10 years, or more frequently if a fecal occult blood test was positive. Additionally, changes in bowel habits or persistent irritable bowel symptoms (unresponsive to medication) could prompt the physician to recommend a colonoscopy on a case-by-case basis. The study included only patients with available biopsy specimens stained with hematoxylin–eosin (see below). The main exclusion criterion for transplant patients was advanced chronic kidney disease (stage IV–V) [25].

A control group was selected, consisting of non-transplanted patients who underwent biopsy of the lower gastrointestinal tract per protocol or within a diagnostic path for gastrointestinal symptoms; these patients were age-, creatinine- and gender-matched with KTRs.

Exclusion criteria for control subjects were administration of immunosuppressants, including the use of anti-rejection drugs, genetic predisposition to tumors, and chronic interstitial diseases. We considered the pathologic region and healthy nearby tissue for every biopsy separately. The clinical characteristics of the two groups are reported in Table 1.

Intestinal tissues were fixed in formalin, embedded in paraffin, and cut into 5-micron thick sections. Paraffin sections mounted on slides were then deparaffinized in Xylene, rehydrated through decreasing concentrations of alcohol and then stained with hematoxylin and eosin. Slides were then dehydrated in alcohol, cleared in xylene and coverslipped in Permount. The sections were then thoroughly digitized.

Images of the biopsies were digitally acquired with a microscope (objective 20×). Several images were acquired to cover the entire region of the histological sample (in some cases up to 1 mm), taking care to have some overlap between the images. Subsequently, the entire section (whole slide images) was reconstructed using the Image Composite Editor 2.0.3.0 64bit by Microsoft. The reconstructed wide fields were then quantitatively analyzed using the ImageJ image analysis software, using the algorithm described below. Two expert pathologists performed histological diagnoses.

A semi-automatic quantitative analysis system was implemented based on manually identifying the position of at least 10 intestinal glands per image masked manually, and a custom ImageJ script. Briefly, RGB images were first split into monochrome channels, and only the green channel was used for further analysis. After background subtraction (rolling algorithm) and automatic local contrast enhancement (CLAHE algorithm), a local threshold algorithm was used to identify all the nuclei in the image automatically (method Sauvola). A watershed filter was then applied to separate clustered nuclei. On these images, the “Analyze particles” command was used to identify the morphology of the gland lumen (area, centroid, shape, perimeter). Subsequently, the perimeter of the mask of each gland lumen was expanded to cover the nuclei of the epithelial cells, and the resulting ribbon covering the epithelial cells of each gland was analyzed first using the “get Profile” command and then linearized (to have a polar-like coordinate system) and each nucleus identified and measured using the “Analyze particles” command. The same method was then used to identify nuclei outside each ring of epithelial nuclei containing the interstitial peri-glandular cells.

For each nucleus, various morphometric parameters were retrieved, such as the size and position, diameter, circularity, density of nuclei per unit length of perimeter, and the number of nuclei. Similar parameters were also obtained for the surrounding nuclei to analyze the microenvironment. Therefore, an average of 50,000 variables were available for each image. To validate the method, two independent pathologists evaluated the intensity of inflammatory cells around glands using a binary scoring system: 0 (normal histology) or 1 (significant inflammation). To further characterize the epithelial layer, we measured the number of mitoses (cell division events) within the gland epithelium. We first defined morphological parameters (optical density, area, solidity, circularity, aspect ratio) of 50 mitoses and 50 non-mitotic epithelial cells identified by an expert observer. The resulting optimal parameters (solidity > 0.8, area < 15, that is small, highly compact nuclei) were then used to develop an automated identification algorithm. The algorithm's reliability was tested by comparing its results to manual counts performed by an independent observer on a larger set of images (521 mitotic events from 20 images). This comparison revealed a significant correlation between the automated and manual measurements (Pearson’s coefficient 0.51, p = 0.02), demonstrating the algorithm's effectiveness.

To further characterize the interstitial cell population and ensure they primarily represent inflammatory cells, we analyzed cells with elongated nuclei, indicative of fibroblasts and endothelial cells. We developed an automated method to identify these cells based on their elongated shape (circularity less than 0.2) compared to other interstitial cells. The accuracy of this method was validated by comparing automated counts to manual counts of endothelial/fibroblast cells (100 fibroblasts/endothelial cells).

Figure 1 shows a flow chart of the sequential steps of the methodology.

Automated image analysis pipeline for colon biopsy assessment. The figure depicts a comprehensive pipeline for analyzing images obtained from gut biopsies. The process starts with image acquisition of the entire section (whole slide imaging). Following acquisition, the pipeline progresses through several steps: (i) pre-processing (grayscale conversion, noise reduction), (ii) segmentation of nuclei to separate epithelial and interstitial cells from the background, (iii) feature extraction for each nucleus (size, shape, staining intensity, spatial distribution of gray levels), (iv) classification: based on the extracted features, algorithms can classify different cell types (epithelial cells, mitoses, endothelial/fibroblast cells, interstitial cells) and quantify their abundance in the tissue

Whole slide imaging data allowed the analysis of each individual nucleus comprising every single gland. Since more than 1700 glands were analyzed, the total number of nuclei analyzed was greater than 60,000, each one characterized by several microscopy parameters. The resulting large dataset was analyzed using R programming environment.

Using R Programming Environment, we segregated parameters from this large dataset, such as Optical Density, area of a nuclear solidity (particle divided by its convex hull area, a measurement of the heterochromatin that is lacking optically empty regions in the nuclei), cell density (number of cells per unit length along the perimeter of the glands) in the glands or interstitial space, and spatial entropy of nuclei (a measurement of the spatial disorganization or disorder of cells). The R script to analyze the ImageJ-produced files is reported in Supplementary Material 2. The obtained data was then analyzed along with Clinical data in SPSS to study statistical significance.

To validate the semi-automatic measurement of infiltrating interstitial cells, we compared the data with a manual binary score assigned to the images by two independent observers. This binary score assessed the presence or absence of inflammation. However, only images where both observers agreed on the degree of inflammation (low or high) were included in the comparison. Images with discordant scores (observers disagreeing) were excluded to ensure reliable validation data. A logistic regression test with manually scored inflammation as the independent variable and density of nuclei in the interstitium (calculated with ImageJ) as a covariate was then used to test the predictive value of the ImageJ scores. We also analyzed the specificity and sensitivity of the automatic measure of interstitial cells for identifying inflammation using ROC curves.

Correlation of morphometric variables between interstitial cells and epithelial cells was tested using Pearson’s correlation coefficient for each gland to find correlations between morphological parameters of the epithelium and the interstitium for each gland.

Differences between the KTRs and the non-KTR control group regarding the clinical variables were tested using Student’s t-test without assuming homogeneity of variance (Welch’s method). To verify the effect of immunosuppressive therapy, we performed separate two-way ANOVAs on the following continuous dependent variables: (i) density of interstitial cells, (ii) density of epithelial cells, (iii) density of endothelial cells, (iv) density of mitoses in epithelial cells. In all cases, the use of immunosuppressive therapy (KTRs vs non-KTRs) and type of lesion (normal tissue vs cancer) were used as categorical independent factors. Continuous data are reported as mean ± standard deviation. The statistical threshold for significance was set at p < 0.05. The power analysis conducted using G*Power initially estimated a sample size of 84 to achieve 80% power for a medium effect size (medium effect size f = 0.31, rejection rate alfa = 0.05). A post-hoc analysis revealed that the actual achieved power based on our data regarding the t-tests of interstitial cells in KTRs and non-KTRs is 99%.