Federated Learning for Chest X-Ray Classification

The experiment

This was my individual undergraduate thesis. I built an experiment pipeline for multi-label chest X-ray classification and used it to compare how several federated-learning frameworks could support a distributed research workflow.

The project was about building and examining the system around the experiment: preparing the data, registering participants, running training, monitoring progress, and checking the resulting artifacts.

What I built

I handled the project independently, from studying the relevant papers and preparing the data through framework comparison, deployment, dashboard development, and analysis. The work included both the researcher-facing flow and the node-operator flow.

How the system ran

The dataset contained 524 chest X-ray images with targets for 38 radiological findings. I used MedGemma-assisted extraction from images and reports as part of the label-creation workflow for the X-Raydar-CV experiments.

I compared FedBioMed, OpenFL, MedFL, NVFlare, and Flower in simulation. I then implemented FedBioMed across five physical devices and completed 30 experimental runs. The system included dashboards for dataset registration, node discovery and setup, training, and monitoring.

Findings and limitations

I evaluated classification metrics, training time, and the completeness of experiment artifacts. The federated configurations did not improve over the pre-federated baseline in this study. The result is useful as an engineering and research finding, but it is not a clinical validation and should not be presented as one.

What I carried forward

The project taught me that a distributed ML experiment is more than a model-training loop. Reproducible setup, participant state, observability, and artifact checks all affect whether an experiment can be interpreted. Those concerns now shape how I approach applied AI systems.