This project explores the use of deep learning for automated detection of uterine (endometrial) cancer from histopathology whole slide images (WSIs). We benchmark multiple neural network architectures—including DenseNet, ResNet, EfficientNet, and Vision Transformers (ViT)—and present a robust pipeline for patch extraction, preprocessing, class balancing, and model evaluation.

Histopathology slides from CPTAC-UCEC See the Project Report for detailed data description.

• Comprehensive Data Pipeline:

  • Patch extraction from WSIs
  • Quality filtering and tissue detection
  • Data augmentation and class balancing

• Model Benchmarking:

  • DenseNet-121
  • ResNet-34
  • EfficientNet-B0
  • Vision Transformer (ViT)

• Performance Highlights:

  • DenseNet and ResNet achieve >99% accuracy
  • CNNs outperform ViT and EfficientNet on limited medical data

• Reproducible Training and Evaluation:

  • Weighted loss for class imbalance
  • Mixed-precision training
  • ROC, PR curves, and confusion matrix analysis

1. Clone the repository:

  git clone https://github.com/krhash/vit-ucec-detection.git
  cd vit-ucec-detection

2. Install dependencies:

  pip install -r requirements.txt

3. Prepare data:

• Download WSIs and labels as described in the data/ folder README.
• Run patch extraction and preprocessing scripts.

4. Train a model:

  python models/train_densenet.py

5. Evaluate results:

• Use scripts in notebooks/ for analysis and visualization.

• DenseNet-121: 99.15% accuracy, AUROC = 1.00
• ResNet-34: 99.10% accuracy, AUROC = 1.00
• EfficientNet-B0, ViT: Underperformed due to limited data

See the Project Report for full details.

View Poster PDF

If you use this code or data, please cite:

Sharma, K.S., Bommireddy, N., Gu, Z. (2025).
Pixels to Prognosis: Deep Learning for Uterine Cancer Detection.
https://github.com/krhash/vit-ucec-detection

This project is licensed under the MIT License. See the LICENSE file for details.

Contact:
For questions or collaboration, please open an issue or contact Krushna Sharma.