Table 4.
Diagnostic performance of the conventional ResNet, shape radiomics classifier, and age plus diameter in the prediction of mitotic index.
| Dataset | AUROC (95% CI) | AUPRC (95% CI) | Acc (95% CI) | |
|---|---|---|---|---|
| Conventional ResNet (per slice) | Development set | 0.951 (0.937–0.966) | 0.960 (0.947–0.970) | 89.9 (87.8–91.7) |
| Test set | 0.927 (0.901–0.953) | 0.929 (0.880–0.959) | 88.7 (85.7–91.1) | |
|
Conventional ResNet
(per patient) a |
Development set | 0.889 (0.820–0.958) | 0.871 (0.769–0.946) | 88.9 (80.2–94.0) |
| Test set | 0.880 (0.760–1.000) | 0.918 (0.746–0.979) | 87.1 (71.2–94.9) | |
| Shape radiomics classifier | Development set | 0.677 (0.641–0.712) | 0.665 (0.629–0.709) | 68.0 (64.9–71.0) |
| Test set | 0.754 (0.712–0.797) | 0.851 (0.823–0.876) | 77.2 (73.5–80.6) | |
| Age plus diameter | Development set | 0.698 (0.574–0.803) | 0.761 (0.657–0.852) | 71.6 (61.0–80.3) |
| Test set | 0.659 (0.465–0.853) | 0.824 (0.690–0.915) | 61.3 (43.8–76.3) |
AUROC, area under the receiver operating characteristics curve; AUPRC, area under the precision–recall curve.
a
Since each patient yielded multiple tumor slices, the diagnostic accuracy per patient was calculated from the mean value of the all-predicted probabilities per patient.