Table 1.
Baseline characteristics of biopsy cores from the training data and the six different test datasets used for evaluation of the AI system and the conformal predictor
| ISUP distribution of biopsies in training and test sets | ||||||||
|---|---|---|---|---|---|---|---|---|
| Cancer grade | Training sets | Test sets | ||||||
| Deep neural network training set (n = 6951) | Conformal prediction calibration set (n = 837) | (1) Baseline test set (n = 794) | (2) Imagebase (n = 87) | (3) External scanner (n = 449) | (4) External scanner and external pathology laboratory (n = 330) | (5) External scanner and external pathology laboratory (n = 1220) | (6) Rare prostate tissue morphology (n = 179) | |
| Benign | 3724 (54%) | 471 (56%) | 440 (55%) | 0 (0%) | 91 (20%) | 108 (33%) | 861 (71%) | 109 (61%) |
| ISUP 1 | 1530 (22%) | 176 (21%) | 172 (22%) | 21 (24%) | 183 (41%) | 65 (20%) | 206 (17%) | 51 (28%) |
| ISUP 2 | 539 (8%) | 80 (10%) | 62 (8%) | 32 (37%) | 64 (14%) | 63 (19%) | 61 (5%) | 19 (11%) |
| ISUP 3 | 263 (4%) | 35 (4%) | 31 (4%) | 15 (17%) | 33 (7%) | 49 (15%) | 45 (4%) | 0 (0%) |
| ISUP 4 | 469 (7%) | 51 (6%) | 41 (5%) | 8 (9%) | 47 (10%) | 19 (6%) | 22 (2%) | 0 (0%) |
| ISUP 5 | 426 (6%) | 24 (3%) | 48 (6%) | 11 (13%) | 31 (7%) | 26 (8%) | 25 (2%) | 0 (0%) |
The Imagebase dataset was independently graded by 23 uropathologists (the mode ISUP grade is shown in the table). ISUP: International Society of Urological Pathology. ISUP 1 (Gleason score 3 + 3), ISUP 2 (Gleason score 3 + 4), ISUP 3 (Gleason score 4 + 3), ISUP 4 (Gleason score 4 + 4, 3 + 5, and 5 + 3), ISUP 5 (Gleason score 4 + 5, 5 + 4, and 5 + 5).