Close this search box.

Computational pathology model to assess acute and chronic transformations of the tubulointerstitial compartment in renal allograft biopsies – Scientific Reports

WSI of slides (n = 1021) stained by modified Picrosirius red during routine kidney allograft biopsy diagnosis were used for the training (n = 852) and internal test subsets (n = 172) in this study. One indication (n = 786) or protocol (n = 238) biopsy containing at least 5 glomeruli per unique patient was randomly selected from a database containing a total of 1021 biopsied patients along with their pathology report data (age, gender, Banff score). The biopsy diagnosis was performed at the National Centre of Pathology, Affiliate of Vilnius University Hospital Santaros Klinikos. The study was performed in accordance with relevant guidelines and regulations and was approved by the Vilnius Regional Biomedical Research Ethics Committee of Vilnius University (reference number: 2019/6–1148-637, date 2019–06-18). Patient informed consent was waived by the Vilnius Bioethics Committee according to the International Ethical Guidelines for Health-related Research Involving Humans33. An external test set (n = 150), comprising both indication (n = 75 and protocol (n = 75) biopsies along with patient data (Banff categories and scores from pathology reports, serum creatinine levels at the biopsy and follow-up period) was used as approved by the Institutional Review Board of the University of California, Davis.

Routine formalin-fixed paraffin-embedded tissue techniques were used in the diagnostic routine. Briefly, sections of 2-μm thickness were stained with modified Picrosirius red staining, which combines Picric acid-Sirius Red solution for staining the cell cytoplasm and collagen fibers and Mayer’s Hematoxylin for staining the cell nuclei and basophilic tissue elements, see (Fig. 1A). The staining procedure involves the steps of deparaffinization of the sections, hydration in distilled H2O, 0.1% Picric acid solution (1 h at room temperature), 0.5% Acetic acid solution (rinse 2–3 times), Mayer’s hematoxylin (5 min), running tap H2O (3 min), dehydration with isopropyl alcohol and clarification with xylene. The final results are evaluated by color indicators: red for collagen, light yellow for cell cytoplasm, and blue for cell nuclei. IHC for CD34 was performed using EnVision FLEX high-sensitivity visualization system on Dako Autostainer Link staining system (DAKO, Glostrup, DK). The CD34 monoclonal mouse antibody (clone QBEnd10; DAKO, Glostrup, DK) was applied at a 1:25 dilution for 30 min, followed by EnVision FLEX + Mouse (Linker) for 15 min. Finally, the sections were developed in DAB at 37 °C for 10 min, counterstained with Mayer’s hematoxylin and mounted. The slides were scanned at the time of diagnosis with Aperio AT2 DX with × 20 objective and pixel size of 0.5 μm (Leica Biosystems, Wetzlar, Germany).

Image analysis

DIA of the WSI was performed using HALO version 3.5.3577 and HaloAI version 3.5.3577 (Indica Labs, Corrales, NM). The Densely Connected Convolutional Networks architecture implemented in HaloAI support sampling network input patches at a specific pixel resolution ( μm) and filtering the segmented objects/regions by a minimum area ( μm2). By using individual classifiers these parameters can be optimized when training classifiers for different kidney histology compartments.

A classifier for renal cortex versus medulla (Fig. 1B and C) and a set of independent one-versus-all classifiers were developed to segment and classify arteries, arterioles and veins (Fig. 1D), peritubular capillaries (PTC) (Fig. 1E), glomeruli (Fig. 1F), tubules (Fig. 1D). By pilot experiments, the following pixel resolution and minimum areas were chosen: healthy and sclerosed glomeruli (1  μm/px, 500  μm2), PTC (0.25  μm/px,40  μm2), tubules (0.25  μm/px, 200  μm2), arteries and veins (2  μm/px, 1000  μm2), cortex and medulla (4  μm/px, 1000  μm2). For the PTC classifier, images of slides re-stained by immunohistochemistry for CD34 (endothelial biomarker) were used to produce an initial set of annotations. To train all the classifiers the “human-in-the-loop machine learning approach”34 was used to generate high numbers of annotations. The annotations were produced or corrected by an image analysis researcher (RA) under guidance of an experienced nephropathologist (AL). Overall, 16,296 annotations were generated in 200 WSI. Segmentation validation definitions and metrics are presented in Supplementary Table S3 and example of PTC validation in Supplementary Figure .

The segmentation output masks produced by the individual classifiers for glomeruli, tubules, peritubular capillaries, arteries, arterioles and veins, were subsequently compiled into a single mask (Fig. S21H). This enabled review, quantification and intersections regions of class overlaps for quality assurance of the segmentation and classification. Examples of compiled classifiers and overlaps are presented in Supplementary Figure S1. For further analyses, overlaps were resolved by preset prioritization rules to correct for typical errors. Most commonly, PTC were “found” in the area of glomeruli and were removed from further analyses by giving a higher priority to glomeruli and other tissue classes. After computational removal of the areas of glomeruli, PTC, tubules, arteries, and veins, the remaining cortical tissue area was presumed to represent the interstitium. This “residual” tissue was further subjected to feature extraction by a classifier to segment red fiber collagen, white–pink non-fiber matrix, and blue nuclei evaluated (Fig. 1I). Tubular compartments (nuclei, cytoplasm, lumen) were further assessed using area quantification (AQ) based on relevant color assessment, blue, light grey and white, respectively, (Fig. 1J).

Sampling criteria

In addition to the initial tissue sampling criterion (> 5 glomeruli by pathology report), we applied a criterion of at least 4 square millimeters of cortical tubulointerstitial area (CTA), after excluding the glomeruli and arteries. This ensured that the tubulointerstitial compartment was well-represented for further analyses independent of potentially crowded glomeruli in some biopsies. In the external set, a criterion of at least 3 square millimeters of CTA was applied due to relatively small amount of tissue in the sections. After implementing this criterion along with the requirement for a minimum of 5 glomeruli, the dataset was reduced to 94 cases from the initial 150.

Feature extraction

For a comprehensive morphometric representation of the tubulointerstitial compartments of renal cortex, we were able to compute 56 compartment-specific features. To avoid an impact of highly variable tissue sample size, relative area indicators of the renal cortex compartments were selected that represent a Cortical Tubulointerstitial Area (CTA) defined as a global sum of the areas of tubular lumen (tLumen%), tubular cells (tCell%), peritubular capillaries (ptc%), and interstitium represented by fibrillary collagen (iColl%) and non-fibrillary matrix (iMatrix%). In addition, 2 global indicators were selected to characterize tubular morphology: density of cell nuclei in tubules (tCellden) and mean area of tubules (tSize). The variable definitions are presented in Table 2.

Table 2 Detailed list of variables, along with their mathematical formulas and definitions, used in the quantitative analysis of renal cortical tissue.

Statistical analyses were performed using RStudio (2022.12.0). To enable comparison of features with different scales, the data were converted into z-scores. Variables were checked for normality of distribution by Shapiro–Wilk statistic; since the data was non-normally distributed, Dunn’s test for pairwise multiple comparisons and Kruskal–Wallis test for more than two groups comparisons were used. To explore feature relationships and select the most relevant ones, PCA was used. The transformation defined by the training PCA model was applied to the testing datasets. In other words, testing set PCA scores represent the data in the reduced dimensional space defined by the principal components derived during the PCA fitting process on the training dataset. K-means clustering was employed to explore potential classifications of the biopsy cases, testing a range of cluster numbers to identify the most meaningful division. The prediction function of K-means, obtained in the training set, was applied to the test set to evaluate clustering reproducibility. Pairwise correlation analysis was performed for all features within each individual cluster. Reciprocal Serum Creatinine slope (rSCs) was calculated individually for each patient using linear regression on at least 4 consecutive (Reciprocal Serum Creatinine (rSC) measurements against the time after the biopsy (n = 44).