Table 4 Distribution of the imaging dataset based on inter-observer and per-annotator agreement levels.

From: GastroHUN an Endoscopy Dataset of Complete Systematic Screening Protocol for the Stomach

Image Dataset Distribution

Strategy

Training label

Team

Train

Valid

Test

Consensus

All

A & B

3,722

793

803

Triple

A & B

5,228

1,103

803

FG

A

4,244

918

803

G

B

5,028

1,078

803

FG1 - G1

A & B

4,940

1,064

803

FG1 - G2

A & B

4,811

988

803

FG2 - G1

A & B

4,553

982

803

FG2 - G2

A & B

4,528

953

803

Annotator

FG(1,2) - G(1,2)

A & B

6,165

1,316

803

Patients

270

58

59

Percentage

70%

15%

15%

  1. The table details the data splits, with the test set held constant across all approaches. “FG” refers to Fellow Gastroenterologists (Team A), and “G” to Gastroenterologists (Team B).