Table 2 Results of the comparison between local and FFL-based training of VinDr-CXR dataset with overlapping labels for different training set sizes, tested on the VinDr-CXR benchmark.

From: Collaborative training of medical artificial intelligence models with non-uniform labels

 

Local VinDr 2K

FFL VinDr 2K

Local VinDr 5K

FFL VinDr 5K

Local VinDr 15K

FFL VinDr 15K

AUROC

0.77 ± 0.08

0.78 ± 0.06

0.79 ± 0.07

0.82 ± 0.05

0.83 ± 0.09

0.84 ± 0.05

P-value

0.340

0.010

0.180

  1. Average AUROC values over no finding, aortic enlargement, pleural thickening, cardiomegaly, pleural effusion, pneumothorax, and atelectasis. The FFL was performed in combination with UKA-CXR dataset of n = 122,294 images with 7 other labels including cardiomegaly, pleural effusion right, pleural effusion left, pneumonic infiltrates right, pneumonic infiltrates left, atelectasis right, and atelectasis left.
Back to article page