UNI-HoverNet: nuclei segmentation and classification across diverse tissue sections based on the UNI foundation model

In this study, we propose UNI-HoverNet, a deep learning model for precise nuclei segmentation and classification across diverse histopathological tissue sections. Our approach integrates the UNI foundation model into the HoverNet encoder, combining the local feature extraction capabilities of CNNs with the global contextual modeling of vision transformers. Additionally, we enhance the HoverNet decoder by incorporating Squeeze-and-Excitation (SE) modules and utilizing skip connections to effectively fuse multi-scale features. This design improves feature utilization while preserving fine-grained structural details, enabling more accurate and efficient nuclei detection. Experiments on the PanNuke public dataset demonstrate that UNI-Hovernet achieves an average multi-class Panoptic Quality (mPQ) of 0.4733, with an average F1-score of approximately 0.53 across five nuclei types. It significantly improves Panoptic Quality (PQ) for inflammatory, connective, and dead nuclei, with increases of approximately 3%, 4%, and 11%, respectively, compared to HoverNet.Clinical relevance- Accurate nuclei segmentation and classification in diverse tissue sections facilitate the analysis of morphometric changes and spatial distribution abnormalities of cell nuclei. These insights can elucidate mechanisms driving tumor inhibition or progression, supporting pathologists in identifying therapeutic strategies or establishing foundations for further research.