NucFuseRank: Dataset Fusion and Performance Ranking for Nuclei Instance Segmentation

This repository contains the implementation details of the NuFuseRank paper.

Two state-of-the-art models (HoVerNeXt [0] and CellViT [1]) were used in this study.

Running Experiments

1. Download the data from the kaggle link below.

2. For HoVerNeXt:

   • Models/HoverNext/run_hovernext.py runs single-dataset training experiments.
   • Models/HoverNext/run_hovernext_all.py runs K-best dataset training experiments.

3. For CellViT:

   • Models/CellVit/CellViT/cell_segmentation/run_cellvit.py runs single-dataset training experiments.
   • Models/CellVit/CellViT/cell_segmentation/run_cellvit_all.py runs K-best dataset training experiments.

4. Evaluate.py and EvaluateAll.py measure model performance on the NucFuse test set.

Note

You need to update the data and weight directory paths according to your local system configuration.

NucFuseRank paper:

In this paper, we evaluated publicly available H&E-stained datasets using two state-of-the-art models: CellViT and HoVerNeXt. We also introduced a fused dataset constructed from these datasets.

Ranking of datasets based on PQ metric.

Dataset	Model	PQ	AJI	Dice	Precision	Recall	PQ Rank	Mean Rank
PCNS	HoVerNeXt	55.66	58.31	68.81	72.64	76.82	2	1
	CellViT	55.14	57.60	68.20	73.21	74.67	1
PUMA	HoVerNeXt	55.75	57.83	71.57	78.62	70.83	1	2
	CellViT	54.74	57.10	72.05	77.89	69.62	3
MoNuSeg	HoVerNeXt	53.38	56.08	68.45	71.05	73.72	3	3
	CellViT	55.11	57.55	69.43	75.09	72.26	2
the th	CPM17	HoVerNeXt	51.86	54.49	65.38	68.05	71.81	4
	CellViT	51.58	54.36	66.18	70.36	70.01	4
TNBC	HoVerNeXt	46.39	48.20	65.94	65.82	62.44	7	5
	CellViT	48.67	52.76	64.85	65.13	68.85	5
NuInsSeg	HoVerNeXt	44.92	44.94	68.04	77.01	55.66	8	6
	CellViT	48.54	50.38	66.06	71.77	63.35	6
CoNSeP	HoVerNeXt	46.41	50.64	63.44	61.02	68.24	6	7
	CellViT	44.97	50.20	62.49	57.05	67.51	7
MoNuSac	HoVerNeXt	47.28	47.28	65.73	77.70	55.34	5	8
	CellViT	40.40	40.40	61.39	71.44	45.84	9
DSB	HoVerNeXt	43.83	45.99	60.13	60.57	61.97	9	9
	CellViT	38.44	44.24	56.53	49.26	59.72	10
CryoNuSeg	HoVerNeXt	35.57	36.57	60.62	60.31	48.16	10	10
	CellViT	40.82	45.49	60.08	55.90	59.91	8
Mean	HoVerNeXt	48.10	50.03	65.81	69.27	64.49
	CellViT	47.84	51.00	64.72	66.71	65.17

NucFuse dataset:

NuFuse dataset is available on kaggle

Acknowledgements

This project has been conducted through a joint WWTF-funded project (Grant ID: 10.47379/LS23006) between the Medical University of Vienna and Danube Private University.

Citation

Our paper preprint is available on arXiv: https://arxiv.org/abs/2601.20104

BibTex entry:

@article{torbati2026nucfuserank,
  title={NucFuseRank: Dataset Fusion and Performance Ranking for Nuclei Instance Segmentation},
  author={Torbati, Nima and Meshcheryakova, Anastasia and Woitek, Ramona and Hatamikia, Sepideh and Mechtcheriakova, Diana and Mahbod, Amirreza},
  journal={arXiv preprint arXiv:2601.20104},
  year={2026}
}

References

• [0] HoVerNeXt
  Paper: https://proceedings.mlr.press/v250/baumann24a.html Code: https://github.com/digitalpathologybern/hover_next_train

• [1] CellViT
  Paper: https://arxiv.org/abs/2306.15350
  Code: https://github.com/TIO-IKIM/CellViT