COIN: Confidence Score-Guided Distillation for Annotation-Free Cell Segmentation

This repository provides the official implementation of COIN, a fully annotation-free framework for cell instance segmentation.
COIN boosts existing UCIS models (e.g., SSA [50]) by recovering error-free instances and distilling them into highly confident pseudo labels—without any image-level annotations or ground-truth masks.

📄 Paper: COIN: Confidence Score-Guided Distillation for Annotation-Free Cell Segmentation
🔗 Project Page: https://shjo-april.github.io/COIN

📰 News

• 📅 ICCV 2025: Our paper has been officially accepted to the IEEE/CVF International Conference on Computer Vision (ICCV) 2025, to be held in October 2025.
• 💻 Coming Soon: We're currently refactoring the codebase for cleaner modularity and improved reproducibility.
• 🚀 Hugging Face Demo: A live demo and pre-trained model checkpoints will soon be released via 🤗 Hugging Face Hub.

📌 Citation

If you use this code or find it helpful, please consider citing:

@InProceedings{jo2025coin,
      title={COIN: Confidence Score-Guided Distillation for Annotation-Free Cell Segmentation}, 
      author={Sanghyun Jo and Seo Jin Lee and Seungwoo Lee and Seohyung Hong and Hyungseok Seo and Kyungsu Kim},
      booktitle={Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV)},
      year={2025}
}

📦 Installation

Set up the environment using:

python3 -m venv venv
source ./venv/bin/activate

python3 -m pip install -r requirements.txt
python3 -m pip install -U "ray[default]" # parallel inference (infer_masks.py)
python3 -m pip install git+https://github.com/lucasb-eyer/pydensecrf.git # for CRF

python3 -m pip install torch==1.13.1+cu117 torchvision==0.14.1+cu117 torchaudio==0.13.1 --extra-index-url https://download.pytorch.org/whl/cu117
python3 -m pip install numpy==1.26.4

📁 Datasets

We provide all six datasets used in COIN experiments, available under the ./data/ directory:

• MoNuSeg
• TNBC
• CPM-17
• CryoNuSeg
• PanNuke
• BRCA

Each dataset is organized using a consistent folder structure, allowing for seamless integration into the pipeline.

📦 Dataset Directory Structure

./data/
├── BRCA/
│   ├── train/
│   │   ├── image/
│   │   └── mask/
│   └── test/
│       ├── image/
│       └── mask/
├── CPM-17/
│   ├── train/
│   │   ├── image/
│   │   └── mask/
│   └── test/
│       ├── image/
│       └── mask/
├── CryoNuSeg/
│   ├── train/
│   │   ├── image/
│   │   └── mask/
│   └── test/
│       ├── image/
│       └── mask/
├── MoNuSeg/
│   ├── train/
│   │   ├── image/
│   │   └── mask/
│   └── test/
│       ├── image/
│       └── mask/
├── PanNuke/
│   ├── Fold1/       # Used as part of training
│   │   ├── image/
│   │   └── mask/
│   ├── Fold2/       # Used as part of training
│   │   ├── image/
│   │   └── mask/
│   └── Fold3/       # Used as test set
│       ├── image/
│       └── mask/
└── TNBC/
    ├── train/
    │   ├── image/
    │   └── mask/
    └── test/
        ├── image/
        └── mask/

All datasets used in this repository were restructured into a consistent folder format to support easy plug-and-play training and evaluation:

./data/{dataset}/{domain}/image/
./data/{dataset}/{domain}/mask/

• For most datasets, domain is train/ or test/
• For PanNuke, we follow a 3-fold split:
  • Fold1 + Fold2 are used for training
  • Fold3 is used for testing

Each dataset is redistributed under its original license, with full attribution to the source authors.
You can find detailed license information and links in LICENSES.md.

⚠️ Note: The CPM-17 dataset is not redistributed here due to unclear licensing. Please download it manually from this repository.

🧩 Step 1: Pixel-level Cell Propagation

# Extract patch-wise USS features from histopathology images using a backbone (e.g., MAE with ViT-B).
# These features are later used to refine the coarse cell masks from a UCIS model.
python3 step1_extract_features_uss.py --uss MAE --backbone ViT-B --data MoNuSeg --domain train --zoom 6 --scales 1.0 --hflip

# Refine initial cell masks using USS features and optimal transport (OT).
# This improves foreground–background separation and ensures the presence of error-free instances.
python3 step1_refine_masks_with_OT.py --uss "./temp/MoNuSeg_MAE@ViT-B/train_{}@hflip/" --scales 6 --ot \
--pred "./submissions/SSA@MoNuSeg@best/train_instance/" --mask "./submissions/Ours+SSA@MoNuSeg@Step1/train/" --accumulate_centroids

🧠 Step 2: Instance-level Confidence Scoring

# Convert refined semantic masks from Step 1 into instance-level masks using connected component labeling and watershed.
# These instance masks will be evaluated for confidence via consistency with SAM-generated masks.
python3 step2_generate_instance_masks.py --data MoNuSeg --domain train --tag "Ours+SSA@MoNuSeg@Step1"

# Score each instance by comparing it with a SAM-generated mask prompted at its center point.
# Instances with high consistency (IoU > mean + std) are retained as confident masks for self-distillation.
python3 step2_score_instances_with_SAM.py --pred "./submissions/Ours+SSA@MoNuSeg@Step1/train_instance/" --mask "./submissions/Ours+SSA@MoNuSeg@Step1/train_sam/" --strategy mean+std 

🔁 Step 3: Confidence Score-guided Recursive Self-distillation

# Train the model using confident pseudo masks from Step 2 (binary masks only; no edge decoder is used).
# This step performs recursive self-distillation to refine predictions and expand confidence progressively.
python3 step3_train_self_distillation.py --gpus 0 --mask "./submissions/Ours+SSA@MoNuSeg@Step1/train_sam/" --tag "Ours+SSA@MoNuSeg@Self-Distillation" --nesterov --scales 100

# Run inference on the test set using the self-distilled model to generate final semantic segmentation predictions.
python3 infer_masks.py --data MoNuSeg --gpus 0 --domain test --tag "Ours+SSA@MoNuSeg@Self-Distillation" --pred "./submissions/Ours+SSA@MoNuSeg@Self-Distillation/test/" --checkpoint last

# Convert the predicted semantic masks from the test set into instance-level outputs using the same semantic-to-instance method from Step 2.
# Note: We reuse watershed-based instance generation instead of training a separate edge decoder.
python3 step2_generate_instance_masks.py --data MoNuSeg --domain test --tag "Ours+SSA@MoNuSeg@Self-Distillation"

🧪 Experiments

After completing all steps, evaluate segmentation quality on the MoNuSeg test set:

▶️ Baseline (SSA only)

python evaluate.py --data MoNuSeg --domain test --tag "SSA@MoNuSeg@best"

Metric	Value
AJI	0.259
PQ	0.185
IoU	0.618
Dice	0.575

---

✅ COIN (Ours + SSA)

python evaluate.py --data MoNuSeg --domain test --tag "Ours+SSA@MoNuSeg"

Metric	Value
AJI	0.580 (+0.321)
PQ	0.536 (+0.351)
IoU	0.776 (+0.158)
Dice	0.794 (+0.219)

For benchmarks/leaderboards, use the corresponding {Model}@{TAG}@{domain}.json. Because COIN uses a consistent schema across datasets, you can evaluate and submit other datasets by simply changing the tag, e.g., Ours+SSA@BRCA@test.json, PSM@CPM-17@test.json, etc.

submissions/
├─ Ours+SSA@MoNuSeg/
│  ├─ test/                 # SSA+COIN: semantic
│  └─ test_instance/        # SSA+COIN: instance
├─ SSA@MoNuSeg/
│  ├─ test/                 # SSA: semantic 
│  └─ test_instance/        # SSA: instance
├─ Ours+SSA@MoNuSeg@test.json
└─ SSA@MoNuSeg@best@test.json

🔎 Summary:
COIN more than doubles the AJI score of the baseline and significantly improves all key metrics—without using any annotations.

🙋‍♀️ Contact

For questions or feedback, feel free to reach out via GitHub Issues or email:

• shjo.april [at] gmail.com