Table 3 Correspondence of chest radiography labels from the RSNA challenge9, NIH3, and NoisyCXR datasets.

From: Active label cleaning for improved dataset quality under resource constraints

 

RSNA labels9

 

Pneumonia-like

No pneumonia-like

 

opacity

opacity

NIH labels3:

  

Pneumonia

367

441

Consolidation/infiltration (not pneumonia)

3,988

8551

Other diseases only

1101

5567

No finding

556

6113

NoisyCXR labels:

  

Pneumonia-like opacity

3956

1296

No pneumonia-like opacity

2056

19,376

Total

6012

20,672

  1. The disagreements between two label sets are highlighted in bold font. Fields in italic indicate an unclear agreement between both sets of labels. NoisyCXR labels are collected from the original NIH labels where possible, and the remainder are uniformly sampled (10%) from the “Consolidation/infiltration” category to increase the noise rate further in the experiments.
Back to article page