Fig. 3.

Results of training on local data only vs. training using FL and SHEFL. Training neural networks on single-site datasets results in inferior performance as compared to FL and SHEFL. A neural network was trained to detect MSI on data from the Epi700, the DACHS and the TCGA cohorts respectively as well as on all three datasets using FL and SHEFL. The resulting networks were then tested on the QUASAR (A) and the YCR-BCIP (B) cohorts demonstrating superior performance of FL and SHEFL. Similarly, tumor segmentation in MRI data was trained on data from five different sites as well as on all data using FL and SHEFL. The resulting neural networks were then tested on an independent held-out test set and demonstrated improved performance (C). Computational overhead was almost negligible (red: overhead for FL, yellow: additional overhead for encryption) as compared to training time needed for backpropagation (blue) (D). .