Figure 2
From: Collaborative training of medical artificial intelligence models with non-uniform labels
Comparison between flexible federated learning (FFL)-based training and local training of classification models. (A) FFL-based training on UKA-CXR data (n = 122,294, labels: pleural effusion right and pneumonic infiltrates right) and on VinDr-CXR data (on 2K, 5K and 15K images, labels: cardiomegaly and pleural effusion) if there is no overlap between labels. Performance tested on an independent VinDr-CXR test set. (B) Same setup as in (A), but training is performed with partially overlapping labels on UKA-CXR (n = 122,294, labels: cardiomegaly, pleural effusion right, pleural effusion left, pneumonic infiltrates right, pneumonic infiltrates left, atelectasis right, and atelectasis left) and on VinDr-CXR (on 2K, 5K and 15K images, labels: over no finding, aortic enlargement, pleural thickening, cardiomegaly, pleural effusion, pneumothorax, and atelectasis). (C) FFL-based training on five different datasets (VinDr-CXR, n = 15,000; ChestX-ray14, n = 86,524; CheXpert, n = 128,356; MIMIC-CXR, n = 210,652; and UKA-CXR, n = 122,294). Testing is performed on the respective held-out test data.
