COSAS – Cross-Scanner Adenocarcinoma Segmentation

A top-10 solution in the MICCAI 2024 COSAS challenge, now published as a full paper at MIE 2025

---

Table of contents

1. About the project
2. Method overview
3. Quick start
4. Results
5. Citations

---

About the project

COSAS explores domain-shift–robust histopathology by jointly learning tumour segmentation and stain separation.
Our joint multi-task U-Net:

• isolates stain matrix & density to handle colour variation,
• boosts mean Dice to 0.898 and IoU to 0.816 on internal validation,
• generalises to six unseen scanners with mean Dice/IoU 0.792. :contentReference[oaicite:1]{index=1}

For challenge details see the official COSAS grand-challenge page. :contentReference[oaicite:2]{index=2}

---

Method overview

``` Our architecture couples a pretrained EfficientNet-B7 encoder with two decoders: * Stain-matrix decoder – learns stain colour bases. * Stain-density decoder – captures tissue structure. Features are fused for segmentation; training minimises a weighted sum of reconstruction and segmentation losses (α ≈ 0.3). See the full paper for details. :contentReference[oaicite:4]{index=4}

Quick start

Requires Docker ≥ 24 & make.

1. Build an inference image around your trained checkpoint

make build MODEL_PATH=/path/to/model.pth

2. Run a quick smoke-test on sample images

make test_run

To export the container:

make save

Results

Dataset / split Dice IoU COSAS internal (3 scanners, 4-fold CV) 0.898 0.816 COSAS external (6 scanners) 0.792 0.792

The submitted model ranked top-10 on the COSAS final leaderboard (Task 2).

Cite

@article{Kim2025COSAS,
  author    = {Ho Heon Kim and Won Chan Jeong and Youngjin Park and Young Sin Ko},
  title     = {Understanding Stain Separation Improves Cross-Scanner Adenocarcinoma Segmentation with Joint Multi-Task Learning},
  journal   = {Studies in Health Technology and Informatics},
  volume    = {327},
  pages     = {53--57},
  year      = {2025},
  doi       = {10.3233/SHTI250272},
  publisher = {IOS Press}
}