Study
Glomeruli are the small capillary structures that filter blood inside the kidney. Locating and counting them in histological tissue is relevant to the assessment of renal health, but manual examination of whole-slide images is slow. Automated segmentation can support this analysis by marking candidate structures and producing counts that a specialist can review.
In this ESANN 2022 paper, written with Simone Bonechi and Giovanna Maria Dimitri, we compared two established deep-learning approaches for glomeruli segmentation. The broader motivation was to study mouse and rat tissue as an accessible first step toward methods that could later be transferred to human samples, for which public annotated data were limited.
The public dataset contains 88 whole-slide images from mouse and rat tissue, acquired with several staining procedures, and 21,037 annotated glomeruli. Whole slides are too large to feed directly to a segmentation network, so we divided them into 10,329 tiles of 512 × 512 pixels. A 90-degree rotation doubled the prepared set to 20,658 image-mask pairs.
This preprocessing step was an important part of the project. It turns sparse annotations distributed across very large slides into manageable training examples while preserving the local tissue context around each glomerulus.
Evaluation
We trained DeepLab V2 and a MobileNet-based segmentation model using the same procedure: Adam optimization, ten training epochs, and five-fold cross-validation. DeepLab V2 uses atrous convolutions to retain spatial detail across a wider receptive field. MobileNet relies on depthwise separable convolutions and is designed for a lighter computational footprint.
The evaluation uses Dice and Jaccard scores to measure overlap between predicted and annotated masks. We also introduced a counting-oriented measure based on connected components: the absolute difference between the number of glomeruli in the prediction and in the ground-truth mask. This matters because a visually plausible mask can still merge nearby structures or split one glomerulus into several components.
DeepLab V2 obtained an average Dice score of 0.91 ± 0.01, a Jaccard score of 0.80 ± 0.03, and a mean absolute glomeruli-count difference of 0.29 ± 0.12. The MobileNet approach obtained a Dice score of 0.81 ± 0.02, a Jaccard score of 0.80 ± 0.02, and a count difference of 0.18 ± 0.11.
Interpretation
The two models reached the same average Jaccard score, while DeepLab V2 obtained the higher Dice score. MobileNet, however, produced the lower counting error. The metrics therefore describe different aspects of the output: DeepLab gives stronger pixel-level overlap, while the lighter model more closely matches the number of annotated objects in these experiments.
This result is the main reason the comparison is useful. It shows that model selection depends on the task performed after segmentation, not only on a single overlap metric. It also makes MobileNet an interesting candidate for future workflows on constrained hardware, including devices connected to microscopy systems.
The paper appears on pages 333-338 of the ESANN 2022 proceedings, and the accompanying repository contains the implementation developed for the project.