Active Challenge • Open for Submissions

Cancerscout HE tumor prediction challenge

Join this exciting challenge to advance medical AI research through privacy-preserving federated learning.

300+
Cases
2+
Clinical Sites
775+
Data Types
Join the Challenge Explore Clinical Data Download Data

Privacy-Preserving

Evaluate on private data without ever seeing it

Fair Evaluation

Unbiased assessment on held-out test sets

Docker-Based

Simple containerized model submission

Clinical Context

Understanding the population, disease characteristics, and biomarkers in this challenge.

Population

Sample Size: 300 cases

  • women/men
  • CRC positive

Disease

  • Nichtinfiltrierendes intraduktales Karzinom o.n.A.
  • Invasives duktales Karzinom o.n.A.

Tissue

  • Zökum
  • Colon transversum
  • Appendix vermiformis
  • Flexura lienalis coli
  • Flexura hepatica
  • Colon sigmoideum
  • Kolon
  • Colon descendens
  • Colon ascendens

Biomarkers

  • HE Stufen (Immun)

Detailed Clinical Information

Population & Study Design

  • women/men
  • CRC positive

Intended Use

Intended to build and verify classifiers for CRC tumors.

Clinical Endpoints

The ground truth labels and evaluation metrics for this challenge.

Clinical Endpoints

The following clinical endpoints are defined for this challenge:

Tumor positive or negative sdfsdf

Indicates the presence of tumorous cells on the slide.

Evaluation Metrics

Your model will be evaluated using the following metrics:

Balanced Accuracy

DESC
Balanced accuracy is a machine learning metric used for evaluating classification models, especially with imbalanced datasets. It is the average of the sensitivity (true positive rate) and specificity (true negative rate). By averaging these two rates, it ensures that both the positive and negative classes are weighted equally, providing a more balanced assessment of performance than standard accuracy when classes are disproportionately represented.

📊 Leaderboard Ranking: Submissions will be ranked based on the metrics above. Metrics marked as ASC (ascending) favor lower values, while DESC (descending) favor higher values.

Challenge Team

Principal Investigator

A

akeil5

Principal Investigator

University Medical Center Göttingen, Institute of Pathology
CAIMed – Lower Saxony Center for Artificial Intelligence and Causal Methods in Medicine

andreas.keil@med.uni-goettingen.de

Contributors

A

akeil5

Contributor

University Medical Center Göttingen

ORCID PubMed Google Scholar

Datasets

Access comprehensive, curated datasets for developing and evaluating your AI models.

Public Dataset

Available for model development and training. Use this data to build and optimize your AI algorithms.

  • ✓ Freely accessible for participants
  • ✓ Annotated with ground truth labels
  • ✓ Comprehensive metadata included
Download Dataset

Private Dataset

Reserved for unbiased model evaluation. Submit your containerized model to test on this hidden data.

  • ✓ Larger cohort for robust evaluation
  • ✓ Never disclosed to participants
  • ✓ Ensures fair model comparison
Evaluation Only

Dataset Statistics

📂 Public Dataset

eval2

725
Total Files
0
Total Cases
907.08 GB
Total Size

File Types Distribution

image/tiff 725 files (100.0%)

Case Statistics

Files per Case
0
Min
0
Avg
0
Max

🔒 Hidden Dataset (Validation)

eval

1703
Total Files
300
Total Cases
2408.63 GB
Total Size

File Types Distribution

image/tiff 1703 files (100.0%)

Case Statistics

Files per Case
2
Min
6.5
Avg
11
Max
Distribution:
2 files:
1 cases
3 files:
1 cases
4 files:
3 cases
5 files:
49 cases
6 files:
135 cases
7 files:
19 cases
8 files:
82 cases
9 files:
9 cases
11 files:
1 cases
Cases per Contributor
akeil5 300 cases

All Terms in Dataset

UFS 1118
HE 796
S360 585
CD3 455
CD8 452

Evaluate Your AI on Private Data

The Centauron Challenge

While the public dataset is extensive, a larger, private portion remains hidden for unbiased model evaluation. You can test your AI model on this hidden data without ever seeing it through the Centauron Network.

This "blind evaluation" ensures a fair and robust assessment of your model's generalizability and performance, mimicking a real-world clinical validation scenario.

How it Works:

  1. Sign Up on the challenge platform over at https://hub.centauron.io/.
  2. Develop Your Model using the public part of the cohort and your own data.
  3. Containerize Your Model into a Docker image.
  4. Submit Your Docker Image to the Centauron platform.
  5. The network runs your model on the private data and returns the performance metrics to you.

For detailed instructions and documentation, visit the official Centauron website at centauron.net.

Ready to test your model?

Participate in our ongoing challenge and see how your algorithm stacks up against others on a truly independent test set.

Participate in the Challenge