Federated Pneumonia Detection Models

ISEF Category: Biomedical Engineering

Subcategory: Other  ·  Difficulty: Advanced  ·  Setup: University Lab  ·  Time: Full Year

The Hook

Hospitals cannot always pool patient data, even when better models could save lives. That creates a big problem for AI training. Federated learning tries to solve it by sending the model to the data, not the other way around. You can test that idea with a pneumonia detection project and compare privacy against accuracy.

What Is It?

Federated learning is a way to train one model across several devices or sites without moving all the raw data into one place. Think of it like a study group where each person keeps their notes, then shares only their answers. A central server combines those updates and sends back a better model. In healthcare, that matters because patient data is sensitive and often locked inside separate hospitals.

This project uses chest X-ray data split across 3 team laptops to mimic 3 hospitals. You can train a pneumonia detector with federated learning, then compare it with a centralized model that sees all the data at once. You can also test differential privacy, which adds noise to model updates so it becomes harder to infer details about individual records. The tradeoff is the key question, how much accuracy do you lose when you protect privacy?

Why This Is a Good Topic

This is a strong science fair topic because you can test a real engineering tradeoff, privacy versus predictive accuracy. You do not need to invent a new disease model from scratch. You can compare training setups, measure model performance, and analyze how privacy settings change the results. That gives you a clean, data-driven story with clear graphs and a practical health-tech angle.

Research Questions

• How does federated learning accuracy compare with centralized training on partitioned CheXpert subsets?
• What is the effect of non-IID data splits on pneumonia detection performance?
• Does adding differential privacy noise reduce model accuracy more in early rounds or later rounds?
• To what extent does client sampling rate change convergence speed in federated training?
• Which aggregation method gives the best balance of accuracy and privacy for this dataset?
• How does the number of participating laptops affect final AUC and loss?

Basic Materials

• Laptop or desktop computers, 3.
• Python installed on each machine, with Jupyter Notebook.
• Free CheXpert subset or another open chest X-ray dataset, pre-split across clients.
• TensorFlow or PyTorch.
• Flower or FedAvg-style federated learning framework.
• Scikit-learn.
• Matplotlib or Seaborn.
• Shared spreadsheet for tracking rounds, settings, and results.

Advanced Materials

• University or lab workstation access, with stronger GPU support.
• Secure network or isolated test environment for multi-client training.
• Docker for reproducible client and server environments.
• PyTorch Lightning or TensorFlow Federated.
• Differential privacy library such as Opacus or TensorFlow Privacy.
• GPU monitoring tools.
• Version control system such as Git.
• Statistical testing tools in Python, such as SciPy and statsmodels.

Software & Tools

• Python: Runs the training code, data splits, and evaluation scripts.
• Jupyter Notebook: Helps you document experiments, plots, and model comparisons in one place.
• Flower: Orchestrates federated learning across multiple client laptops.
• PyTorch: Trains the pneumonia detection model and supports custom training loops.
• Scikit-learn: Calculates AUC, precision, recall, confusion matrices, and baseline metrics.

Experiment Steps

1. Define the exact question you want to test, such as accuracy versus privacy or IID versus non-IID data splits.
2. Choose one chest X-ray dataset and decide how you will partition it across client laptops to mimic separate hospitals.
3. Build a centralized baseline first so you have a fair reference point for every federated result.
4. Design the federated setup, including the number of clients, the aggregation rule, and the privacy settings you will compare.
5. Plan the evaluation metrics before training starts, including AUC, recall, loss curves, and any privacy budget measure.
6. Prepare controls for data imbalance, repeated runs, and random seeds so your results are not just luck.

Common Pitfalls

• Using a random train-test split after partitioning, which leaks information across simulated hospitals and inflates performance.
• Comparing federated and centralized models with different model architectures, which makes the results unfair.
• Ignoring class imbalance in pneumonia labels, which can make accuracy look good while recall stays weak.
• Treating every client split as IID when real hospital data usually has different label mixes and imaging patterns.
• Reporting privacy budget values without explaining the accuracy tradeoff, which leaves the core result incomplete.

What Makes This Competitive

A competitive version of this project does more than run one federated model. You compare multiple partition patterns, privacy settings, and aggregation rules, then test whether the results stay stable across repeated trials. You can also add a stronger analysis of clinical metrics like recall and false negatives, since missing pneumonia matters more than overall accuracy. If you clearly explain the tradeoffs and back them with careful statistics, your project looks much stronger.

Project Variations

• Use a different chest X-ray dataset, then test whether federated performance changes when image quality or labeling style changes.
• Compare simple FedAvg against a privacy-preserving aggregation method to see which one keeps the best balance of accuracy and privacy.
• Swap the pneumonia task for another binary medical imaging task, such as edema or pleural effusion detection, to test whether the same setup generalizes.

Learn More

• Flower Documentation: Read the official federated learning tutorials and examples on the Flower project site.
• TensorFlow Federated: Explore Google’s open federated learning framework and its tutorials on the TensorFlow site.
• PyTorch Tutorials: Find model-building and training examples on the official PyTorch website.
• PubMed: Search review articles on federated learning, differential privacy, and medical imaging.
• NIH PubMed Central: Read full-text biomedical papers on privacy-preserving machine learning in healthcare.
• CheXpert Paper and Dataset Materials: Search for the CheXpert study paper and related dataset descriptions on arXiv, Stanford pages, or PubMed.