john121319/federated-pneumonia-research ? reverse-engineered prompt

Build me a research project that reproduces a federated learning study for pneumonia detection from chest X ray images.

I want it to take the RSNA pneumonia dataset, split it by patient so the same person never ends up in training, validation, and test at the same time, then simulate a few federated clients with different levels of data imbalance. Train a simple image classifier centrally, with FedAvg, and with FedProx, then compare how they do as the client data gets more non uniform.

Please include the full preprocessing flow for the DICOM images, a cached image output, reproducible runs with fixed seeds, validation only model selection, and a final test step that uses frozen checkpoints and chosen thresholds. Also save the key metrics and figures so I can inspect PR AUC, ROC AUC, F1, balanced accuracy, and how much the client updates drift during training.

If you need to, look up current docs online for any library details.