Kidney Tissue Segmentation
About
Summary
Inflammation and chronic changes in the different tissue structures (e.g., glomeruli, tubuli, interstitium) are major contributors to kidney transplant failure. Kidney transplant biopsy diagnostics is based on the Banff classification system, in which pathologists assess these changes. However, many of these factors have suboptimal reproducibility and the scoring is labor-intensive. To address this, we developed a multi-class segmentation approach that covers all tissue structures relevant for diagnostics.
Our dataset comprises 99 Periodic-acid Schiff (PAS)-stained kidney transplant biopsy slides from two pathology departments. An expert pathologist manually annotated >17,000 structures across eight classes (glomeruli, sclerotic glomeruli, empty Bowman space, proximal tubuli, distal tubuli, atrophic tubuli, capsule, arteries/arterioles, and interstitium). The nnU-Nets achieved a per-class average Dice score of 0.80, providing a reliable solution for all tissue structures relevant for kidney transplant biopsy diagnostics.
Mechanism
Training data
The algorithm targets kidney transplant recipients. It was developed using PAS-stained biopsy slides from patients at multiple centers, encompassing a diverse set of samples. Expert annotations include eight renal structures: glomeruli, sclerotic glomeruli, empty Bowman's space, tubuli, atrophic tubuli, capsule, arteries/arterioles, and interstitium, capturing the full range of histological features relevant to transplant pathology.
Methods
The nnU-Net used for this algorithm was trained using 512*512 px patches at 1.0 µm/px resolution with 5-fold cross-validation, where fold assignments were optimized to match the dataset-wide class pixel distribution. The network was trained for 100 iterations with a batch size of eight, using batch normalization, LeakyReLU activation (slope 0.01), and a set of augmentations including spatial (scaling, rotation, mirroring), color (HSV, gamma), and deformation transforms. During inference, tissue-background segmentation masked non-tissue areas, followed by patch-wise sliding window prediction with half-overlap and Gaussian importance weighting. To better resolve adjacent or overlapping structures, a separate border nnU-Net was trained to segment three classes: structure, border, and interstitium, using the same training configuration as the baseline model. Borders were generated by dilating ground truth masks by eight pixels (four inward, four outward)
Inputs and outputs
The algorithm takes as input a PAS-stained whole-slide image of a kidney transplant biopsy together with a binary tissue mask. The tissue mask covers the regions of the WSI that should be segmented during inference. After inference is complete, the output TIF mask covers the regions of the input mask with labels from 1 to 8.
Interfaces
This algorithm implements all of the following input-output combinations:
Validation and Performance
| nnU-Net | Standalone Dice | With Border Dice |
|---|---|---|
| 0.944 | 0.956 | |
| 0.926 | 0.924 | |
| 0.888 | 0.923 | |
| 0.902 | 0.899 | |
| 0.657 | 0.657 | |
| 0.893 | 0.925 | |
| 0.951 | 0.933 | |
| 0.872 | 0.874 | |
| Weighted average | 0.885 | 0.905 |
| Per-class average | 0.873 | 0.886 |