Table 3 Results of the reader study in the gastroscopy data set.

From: Weakly supervised end-to-end artificial intelligence in gastrointestinal endoscopy

ICD

Diagnosis

Typical

Not typical

Missing

Device

Artifact

B37

Candidiasis

0.68

0.18

0.03

0.07

0.03

K26

Duodenal ulcer

0.60

0.23

0.00

0.15

0.02

K20

Oesophagitis

0.57

0.32

0.02

0.08

0.02

K22

Other diseases of oesophagus

0.55

0.15

0.22

0.05

0.03

C15

Malignant neoplasm of oesophagus

0.50

0.13

0.13

0.18

0.05

K29

Gastritis and duodenitis

0.48

0.23

0.03

0.18

0.07

C16

Malignant neoplasm of stomach

0.48

0.33

0.05

0.13

0.00

K25

Gastric ulcer

0.47

0.33

0.10

0.08

0.02

K92

Other diseases of digestive system

0.47

0.35

0.03

0.15

0.00

D13

Benign neoplasm of other and ill-defined parts of digestive system

0.42

0.32

0.07

0.18

0.02

K21

Gastro-oesophageal reflux disease

0.42

0.38

0.02

0.18

0.00

CXX

Malignant neoplasm of oesophagus or stomach

0.40

0.17

0.12

0.28

0.03

R11

Nausea and vomiting

0.38

0.37

0.08

0.13

0.03

K31

Other diseases of stomach and duodenum

0.37

0.35

0.02

0.22

0.05

D37

Neoplasm of uncertain/unknown behaviour, oral cavity/digestive organs

0.37

0.43

0.02

0.17

0.02

R13

Dysphagia

0.28

0.47

0.12

0.10

0.03

I85

Oesophageal varices

0.27

0.48

0.03

0.20

0.02

D48

Neoplasm of uncertain/unknown behaviour of other/unspecified sites

0.25

0.43

0.07

0.23

0.02

K57

Diverticular disease of intestine

0.20

0.68

0.00

0.10

0.02

R12

Heartburn

0.18

0.65

0.05

0.10

0.02

K91

Postprocedural disorders of digestive system, not elsewhere classified

0.17

0.28

0.10

0.42

0.03

K44

Diaphragmatic hernia

0.12

0.57

0.00

0.30

0.02

R10

Abdominal and pelvic pain

0.10

0.75

0.08

0.05

0.02

  1. For each diagnosis category in the gastroscopy dataset, the AI system selected 60 highly predictive images (2 highest scoring images from 10 highest scoring patients in each of the 3 folds). These images were manually classified into five categories by a trained observer. The fraction of images in each category in every category is shown.