Table 1:
Categorization of a random subset of 1578 true-positive and all 228 false-negative polyp CNN predictions on the test set of 4088 unique polyps, categorized by size/Paris classification. Results obtained via 7-fold cross-validation on the 8641 colonoscopy images. All polyps >3mm are categorized by the Paris classification scheme. The CNN performs equally well at detecting nonpolypoid lesions (II a/b/c) and polypoid polyps (I p/s).
| True Positives | False Negatives | |||
|---|---|---|---|---|
| ≤ 1cm | > 1cm | ≤ 1cm | > 1cm | |
| Dim (≤ 3 mm) | 644 | - | 103 | - |
| I p | 37 | 25 | 8 | 6 |
| I s | 487 | 45 | 68 | 2 |
| II a | 246 | 37 | 36 | 4 |
| II b | 34 | 15 | 1 | 0 |
| II c | 4 | 4 | 0 | 0 |