Table 4.
Potential reasons for incorrect classification of external-test dataset 2 images by the established no-code tool-based deep-learning models.
| Unknown (Difficult Cases Even for Endoscopists) | Multiple Attention or Partial Attention Even Though the Image Was Appropriate | Normal Mucosal Folds or Blood Vessels Recognised as Lesions | Inappropriate Images (Only a Part of the Lesion Can Be Observed) | Inappropriate Images (Multiple Lesions Were Observed in One Image) | Inappropriate Images (Residual Food or a Bubble Was Recognised as a Lesion) | |
|---|---|---|---|---|---|---|
| Advanced colorectal cancers | ||||||
| Incorrectly diagnosed as early colorectal cancers/high-grade dysplasias (n = 10) | 4 | 5 | 1 | |||
| Incorrectly diagnosed as non-neoplasm (n = 1) | 1 | |||||
| Early colorectal cancers/high-grade dysplasias | ||||||
| Incorrectly diagnosed as tubular adenoma (n = 56) | 47 | 9 | ||||
| Incorrectly diagnosed as non-neoplasm (n = 15) | 1 | 14 | ||||
| Incorrectly diagnosed as advanced colorectal cancers (n = 7) | 3 | 4 | ||||
| Tubular adenomas | ||||||
| Incorrectly diagnosed as non-neoplasm (n = 70) | 27 | 35 | 6 | 2 | ||
| Incorrectly diagnosed as early colorectal cancers/high-grade dysplasias (n = 20) | 12 | 5 | 1 | 1 | 1 | |
| Non-neoplasms | ||||||
| Incorrectly diagnosed as tubular adenoma (n = 24) | 3 | 20 | 1 | |||
| Total | 94 (46.3%) | 76 (37.4%) | 27 (13.3%) | 1 (0.5%) | 3 (1.5%) | 2 (1%) |
External-test dataset 2: from Kangdong Sacred Heart Hospital.