HistoMSC: density and topology analysis for AI-based visual annotation of histopathology whole slide images

We introduce an end-to-end framework for the automated visual annotation of histopathology whole slide images. Our method integrates deep learning models to achieve precise localization and classification of cell nuclei with spatial data aggregation to extend classes of sparsely distributed nuclei across the entire slide. We introduce a novel and cost-effective approach to localization, leveraging a U-Net architecture and a ResNet-50 backbone. The performance is boosted through color normalization techniques, helping achieve robustness under color variations resulting from diverse scanners and staining reagents. The framework is complemented by a YOLO detection architecture, augmented with generative methods. We use context patches around each nucleus, fed to various deep architectures, for classification. Sparse nuclei-level annotations are then aggregated using kernel density estimation, followed by color-coding and isocontouring. This reduces visual clutter and provides per-pixel probabilities with respect to pathology taxonomies. Finally, we use Morse-Smale theory to generate abstract annotations, highlighting extrema in the density functions as well as highlighting potential spatial interactions in the form of abstract graphs. Thus, our visualization allows for exploration at scales ranging from individual nuclei to the macro-scale. We tested the effectiveness of our framework in an assessment by six pathologists using various neoplastic cases. Our results demonstrate the robustness and usefulness of the proposed framework in aiding histopathologists in their analysis and interpretation of whole slide images.

@Article{ZAAFGMBPGS25,
  author       = {Zahoor, A. and Al Thelaya, K. and Alzubaidi, M. and Faaiz, J. and Gilal, N. and Mifsud, W. and Boughorbei, S. and Pintore, G. and Gobbetti, E. and Schneider, J.},
  title        = {HistoMSC: density and topology analysis for AI-based visual annotation of histopathology whole slide  images},
  journal      = {Computers in Biology and Medicine},
  volume       = {190},
  pages        = {109991},
  month        = {may},
  year         = {2025},
  publisher    = {Springer},
  keywords     = {visual computing, data-intensive computing},
  doi          = {10.1016/j.compbiomed.2025.109991},
  url          = {https://publications.crs4.it/pubdocs/2025/ZAAFGMBPGS25},
}