UNet-HoVerGNN: structured graph integration into HoVerNet for enhanced nuclei segmentation and classification

We propose UNet-HoVerGNN, a hybrid framework for nuclei instance segmentation and classification that combines CNN-based segmentation with graph-based reasoning. Nuclei centroids are used to construct k-nearest neighbor graphs, where node and edge features are extracted from a pretrained U-Net encoder. A GNN refines classification by capturing spatial and contextual dependencies. Evaluated on MoNuSAC, PanNuke, and CoNSeP, our model achieves strong results: 0.603 AJI, 0.641 F1, 0.622 PQ on MoNuSAC; 0.644 AJI, 0.682 F1, 0.661 PQ on PanNuke; and 0.585 AJI, 0.608 F1, 0.595 PQ on CoNSeP-outperforming or matching state-of-the-art models like HoVerNet and similar graph-based methods. Code is available at https://github.com/YangTuanAnh/UNet-HoVerGNN.