Table 4 Individuals with Bacteriologically Confirmed TB, Missed by Deep Learning Systems at 95% Sensitivity.

From: Using artificial intelligence to read chest radiographs for tuberculosis detection: A multi-site evaluation of the diagnostic accuracy of three deep learning systems

Individual Missed

Senior Radiologist (Nepal)

Junior Radiologist & Residents (Nepal)

Field Radiologist (Cameroon)

Teleradiology Company (Cameroon)

DL reading

CAD4TB score

qXR score

Lunit score

Annotation by a senior pulmonologist

Nepal 1

Normal

Normal

NA

NA

Missed by all 3 DL systems

9

0.1205

0.2330

Normal

Nepal 2

Abnormal

Abnormal

NA

NA

Missed by qXR

65

0.4225

0.9853

Abnormal: may be an azygous lobe (normal variant) but also could be apical TB

Nepal 3

Abnormal

Abnormal

NA

NA

Missed by Lunit

71

0.5223

0.3458

Abnormal maybe old scar: minimal tenting of the right diaphragm

Nepal 4

Normal

Normal

NA

NA

Missed by Lunit and qXR

63

0.1517

0.1318

Non-TB abnormality: elevated right hemidiaphragm

Nepal 5

Normal

Normal

NA

NA

Missed by CAD4TB and Lunit

46

0.4992

0.2728

Normal

Cameroon 1

NA

NA

Normal

Abnormal

Missed by all 3 DL systems

48

0.1259

0.0176

Abnormal in the left mid lung field by the heart- could be TB but not classic

Cameroon 2

NA

NA

Normal

Normal

Missed by CAD4TB

53

0.6695

0.8744

Abnormal: ill defined infiltrate in the right upper lung field

Cameroon 3

NA

NA

Normal

Normal

Missed by CAD4TB and qXR

18

0.2504

0.5518

Normal lung field but cardiomegaly