Table. Detection of Cervical Cell Atypia With the Deep Learning System in Digitized Papanicolaou Tests, Compared With Expert Assessments of Digitized and Physical Slidesa.
| Diagnostic comparison | % (95% CI) | No. (%) | ||||
|---|---|---|---|---|---|---|
| True | False | |||||
| Sensitivity | Specificity | Positive | Negative | Positive | Negative | |
| Digitized-slide cytodiagnosis | ||||||
| General atypia | 95.7 (85.5-99.5) | 84.7 (80.2-88.5) | 45 (12.5) | 266 (73.7) | 48 (13.3) | 2 (0.6) |
| High-grade atypia | 85.7 (67.3-96.0) | 98.5 (96.5-99.5) | 24 (6.6) | 328 (90.9) | 5 (1.4) | 4 (1.1)b |
| Low-grade atypia | 84.2 (60.4-96.6) | 86.0 (81.8-89.5) | 16 (4.4) | 294 (81.4) | 48 (13.3) | 3 (0.8) |
| Glass-slide cytodiagnosis | ||||||
| General atypia | 100.0 (82.4-100.0) | 78.4 (73.6-82.6) | 19 (5.3) | 268 (74.2) | 74 (20.5) | 0 |
| High-grade atypia | 100.0 (47.8-100.0) | 93.3 (90.1-95.6) | 5 (1.4) | 332 (92.0) | 24 (6.6) | 0 |
| Low-grade atypia | 21.4 (4.7-50.8) | 82.4 (78.0-86.3) | 3 (0.8) | 286 (79.2) | 61 (16.9) | 11 (3.0)c |
a
Sensitivity and specificity results from the deep learning system are shown with the associated 95% CIs. Numbers of false-negative, false-positive, true-negative, and true-positive assessments are shown with the corresponding percentage of the total number of slides in the validation series (n = 361).
b
Four slides identified as having high-grade atypia were classified as low-grade atypia by the deep learning system.
c
Eleven slides identified by the local pathologist as low-grade atypia were classified as high-grade atypia by the deep learning system.