RankSEG replaces argmax or fixed thresholds with metric-aware post-processing designed to improve Dice or IoU—no retraining or fine-tuning. Use it with frozen PyTorch segmentation models for multiclass, binary, and multilabel tasks, from natural images to 3D medical scans.

RankSEG on 3D CT · Same model. No retraining.

20-volume mean Dice: 50.30 → 54.87 (+4.57 pp)
Frozen MONAI BTCV Swin UNETR · MSD Pancreas.

Illustrative slice above; white outline = ground truth. Research use only.
Evaluation protocol and example selection.

📰 News

• August 2026 — RankSEG joins the official MONAI Tutorials! Try metric-aware post-processing for 3D medical segmentation. Tutorial · Colab

⚡ Quick Start

pip install -U rankseg

Optional Linux x86-64 GPU acceleration: pip install "rankseg[cuda]" (setup and activation).

For multiclass model logits shaped (batch, classes, *spatial), with at least two classes:

from rankseg import RankSEG

probs = model_logits.softmax(dim=1)
preds = RankSEG(metric="dice")(probs)  # replaces argmax; shape: (batch, *spatial)

For binary/multilabel examples, the functional API, and solver options, see the Getting Started guide.

RMA Dice offers optional memory-saving screening with safe_screening="auto". The default remains False, preserving the original computation path. Opting in may change masks near numerical ties. Details.

Try it online: Colab · Interactive demo

📊 Benchmarks

Same probabilities. No retraining. Selected results compare argmax with RankSEG-RMA using the Dice objective. Scores are percentages; gains are percentage points (pp).

Gains vary by dataset. Full results, metric definitions, runtime, and reproduction commands: rankseg-benchmark. For the algorithm and additional experiments, see our NeurIPS 2025 paper.

🔌 Integrations

Choose your workflow:

🔗 Citation

If you use RankSEG in your research, please cite our papers:

• Dai, B., & Li, C. (2023). RankSEG: A Consistent Ranking-based Framework for Segmentation. Journal of Machine Learning Research, 24(224), 1-50. [link]
• Wang, Z., & Dai, B. (2025). RankSEG-RMA: An Efficient Segmentation Algorithm via Reciprocal Moment Approximation. Advances in Neural Information Processing Systems (NeurIPS 2025). [link]

@article{dai2023rankseg,
  title={RankSEG: A Consistent Ranking-based Framework for Segmentation},
  author={Dai, Ben and Li, Chunlin},
  journal={Journal of Machine Learning Research},
  volume={24},
  number={224},
  pages={1--50},
  url={https://www.jmlr.org/papers/v24/22-0712.html},
  year={2023}
}

@inproceedings{wang2025rankseg,
  title={RankSEG-RMA: An Efficient Segmentation Algorithm via Reciprocal Moment Approximation},
  author={Wang, Zixun and Dai, Ben},
  booktitle={Advances in Neural Information Processing Systems},
  url={https://arxiv.org/abs/2510.15362},
  year={2025}
}