“No Negatives Needed”: Weakly-Supervised Regression for Tumor Detection in Histopathology 🔬

This repository contains the implementation of the method described in the paper: “No negatives needed”: weakly-supervised regression for interpretable tumor detection in whole-slide histopathology images". We propose a novel approach using Multiple Instance Learning (MIL) in a regression setting for tumor detection in whole-slide images. This approach offers solution to train tumor detection models while only having positive cases with coarse annotations. Our method leverages tumor area percentages as continuous targets, used as a proxy for tumor detection.

✨ Key features

• Weakly-Supervised Learning: No need for detailed pixel-level annotations.
• Tumor Area Percentage Regression: Predicts continuous tumor area percentages as a proxy for tumor detection without needing negative cases for training.
• Support of MIL models: Supports various MIL approaches such as mean pooling (MeanPool), attention-based MIL (ABMIL), (CLAM) and (WeSEG).
• Noise Robustness: An analysis of the model's robustness to synthetic noise in tumor percentage labels demonstrated that the models maintain strong performance even when faced with variability and noise in clinical annotations.
• Explainability: Generates attention and logits heatmaps for interpretable AI applications.

🛠️ How to Use

Ensure you have Docker installed and running on your system.

1. Clone this repository.

git clone https://github.com/DIAGNijmegen/tumor-percentage-MIL-regression.git
cd tumor-percentage-MIL-regression

2. Prepare a config file to customize the model training, data paths, optimizer, and other settings. An example of the configuration file is provided in config\default.yaml.

   • The data section in the yaml config file contains all the paths to data and results directories. You can find examples of splits and csv_path under examples. split_dir is the path to the directory containing splits_{n_fold}.csv for each fold, with columns "train", "val", "test" specifying slide IDs for each split. features_dirs is the path(s) to the pre-computed feature vectors. Multiple directories can be specified (e.g. when training on multiple datasets with features stored in different folders). results_dir is the ouput folder. csv_path is the path to a csv file with columns "slide_id" and "label".

3. Build the Docker image

docker build -t tumor-detection .

4. Run the Docker container, mapping your local directories to the container. Adjust the volume mappings to match your data paths in the config file:

docker run -it -v /path/to/data/:/path/to/data/ -v /path/to/config:/path/to/config tumor-detection /bin/bash

5. Once inside the Docker container, run the main script using the appropriate configuration file:

python3 main.py --config-name name_config

📊 Output and interpretability

The model outputs include:

• Predicted tumor percentages (for MeanPool, ABMIL, CLAM).
• Instance-level predictions.
• Attention scores (for ABMIL and CLAM). Instance-level predictions and attention scores can be visualized using heatmaps. Heatmap generation relies on code adapted from CLAM to generate heatmaps. To create heatmaps, use the script located in utils/generate_heatmaps.py, which expects a configuration file similar to config-heatmaps/default.yaml.

🧪 Testing

To test an existing checkpoint or evaluate the model:

• Specify the ckpt_path in the configuration file under the testing section.
• Set only_testing: True to skip training and directly evaluate.

💡 Tips for Best Results

To improve tumor detection performance and interpretability, consider using the amplification technique introduced in our paper. This method was especially effective in:

• Datasets with biopsies or lymph nodes.
• Scenarios with small lesions or tumor-free cases, where differentiating between very small tumor percentages and negative slides is challenging.

The amplification technique involves applying a transformation to the tumor area percentage targets during training. In our experiments, a fifth-root transformation was effective. However, you are encouraged to explore transformations that best suit your dataset.

📜 Additional scripts

• utils/fold_splits.py: Generates 5-fold splits with stratification on continuous targets.
• utils/add_noise.py: Adds uniform noise to training targets to simulate variability in clinical practice.
• utils/generate_heatmaps.py: Creates heatmaps for instance-level predictions and attention scores.

💬 Feedback & Support

Feel free to reach out with questions, suggestions, or issues via the repository's Issues tab or contact the authors directly.

🏆 Acknowledgments

This work builds upon contributions from the following repositories:

• hs2p: Used for patch extraction.
• CLAM: Used for feature extraction, heatmap generation, and CLAM model.
• TCGA Segmenttion: Starting point for developing the WeSEG model.

This project has received funding from the Innovative Medicines Initiative 2 Joint Undertaking under grant agreement No 945358. This Joint Undertaking receives support from the European Union’s Horizon 2020 research and innovation program and EFPIA (www.imi.europe.eu).