CytoSPM

Model weights of CytoSPM: Open-Vocabulary Cytopathology Detection with Structured Prompt Bank (BIBM 2026).

• Paper: arXiv:2609.31314
• Code: github.com/CVIU-CSU/CytoSPM

Files

File	Backbone	Base AP / AP50 / AP75	Novel AP / AP50 / AP75	SHA256
cytospm_convnext_t.pth	ConvNeXt-T	33.54 / 52.20 / 38.16	15.00 / 21.00 / 17.10	bc94215314319f712c731e8f5f24388df29d254cc2029695d32e815efb3aac4d

This is the model reported in the paper (Table III). The numbers above were measured with the released code on an RTX 3090; the paper values (33.55 / 52.20 / 38.17 and 15.00 / 21.00 / 17.20) were measured on an RTX 4090. Base AP is the COCO AP over the 24 base categories of PentaCyto; novel AP is the zero-shot COCO AP over the 9 held-out categories.

The file contains model weights only (state_dict) and loads with torch.load(..., weights_only=True).

Usage

git clone https://github.com/CVIU-CSU/CytoSPM.git && cd CytoSPM
CKPT_FILES=cytospm_convnext_t.pth bash tools/fetch_checkpoints.sh   # downloads and checks SHA256
python tools/test.py configs/cytospm.py checkpoints/cytospm_convnext_t.pth
python tools/test_novel.py configs/cytospm.py checkpoints/cytospm_convnext_t.pth

See the code repository for installation, data preparation and training. The PentaCyto images and annotations are not included.

Citation

@inproceedings{li2026cytospm,
  title     = {CytoSPM: Open-Vocabulary Cytopathology Detection with Structured Prompt Bank},
  author    = {Li, Wenjie and Xu, Zishan and Huang, Jinyang and Nie, Zhengxin and Kan, Shichao and Liang, Yixiong},
  booktitle = {IEEE International Conference on Bioinformatics and Biomedicine (BIBM)},
  year      = {2026}
}