We are looking for a Master’s student to develop and benchmark a novel federated aggregation method for multi-site medical image analysis. The project focuses entirely on algorithm development using public datasets (ImageCAS and BraTS), making it fully self-contained and completable within a six-month full-time thesis.
The core challenge is handling scanner and acquisition heterogeneity across sites without dataset-specific assumptions. Building on established approaches such as FedBN and FedProx, the student will design a heterogeneity-aware aggregation or normalization scheme that generalizes across datasets. The work involves reproducing FedAvg as a baseline, developing the proposed method, running ablations against FedAvg, FedProx, and FedBN, and writing up the findings — all implemented in plain PyTorch.
Requirements: Prior familiarity with federated learning is essential. Given the six-month timeline and the expectation of developing a novel algorithm, students without existing FL experience will not be able to complete the scope of work.