Glomeruli detection and classification in histopathological images using deep learning semantic segmentation.

BACKGROUND AND

objectiveGlomeruli are essential renal structures responsible for blood filtration, and their condition is indicative of various kidney diseases. Accurate detection and classification of glomeruli in histopathological images are critical for early diagnosis and treatment planning. This study presents an automated method for glomeruli segmentation and classification using deep learning-based semantic segmentation, aiming to reduce inter-observer variability and support diagnostic workflows.

methodsWe propose a modified version of the DeepLab v3 + architecture, where we contributed by enhancing it with architectural adaptations for fine-grained tissue analysis, mainly related to the encoder multiscale feature extraction. The model incorporates a refined Atrous Spatial Pyramid Pooling (ASPP) module and uses the Xception backbone for improved feature extraction. A two-phase training strategy was implemented: initial training on a large semi-automatically annotated dataset, followed by fine-tuning on a smaller, manually labeled set. Images were preprocessed using Contrast-Limited Adaptive Histogram Equalization (CLAHE) and divided into overlapping patches to preserve structural details. The model was trained using a combined loss function of categorical focal Cross-Entropy (CE) and Tversky loss, and performance was evaluated using standard metrics such as Intersection over Union (IoU) and Dice Similarity Coefficient (DSC), reported both overall and per class, as well as class-wise recall and misclassification rates.

resultsThe proposed model achieved an overall IoU of 0.7674 and a DSC of 0.8614 on the test set. For non-sclerotic glomeruli, the model reached a DSC of 0.8766 and a sensitivity of 84.90%. For sclerotic glomeruli, it achieved a DSC of 0.6952 and a sensitivity of 59.25%. The model consistently outperformed baseline architectures including U-Net, SegNet, nnU-Net, Swin UNETR and the standard DeepLab v3+, particularly in cases involving challenging histological variations. Performance was highest for Hematoxylin and Eosin-stained images, with a DSC of 0.8531.

conclusionsThe proposed method demonstrates strong potential for supporting renal pathology workflows by providing accurate glomerular segmentation and classification. Although a key limitation is that performance on sclerotic glomeruli remains lower due to class imbalance and stain variability, the approach shows significant improvement over existing models. Future work will explore stain normalization and advanced architectures to enhance segmentation robustness across diverse histological conditions.