Table 2.
Performance of DNN model at points of high sensitivity of detecting advanced neoplasia in colorectal cancer screening
| Screening strategy | Sensitivity, % | Specificity, % | No. of colonoscopy | ACRNs detected, n | NNScope, n | Reduction of NNScope, % | p value (vs. LR) |
|---|---|---|---|---|---|---|---|
| Target sensitivity | 80 | ||||||
| LR | 78.9 | 51.7 | 6,855 | 146 | 47.0 | Reference | Reference |
| DNN | 79.5 | 58.2 | 5,948 | 147 | 40.5 | 13.8 | < 0.001 |
| Target sensitivity | 90 | ||||||
| LR | 89.2 | 26.5 | 10,364 | 165 | 62.8 | Reference | Reference |
| DNN | 89.7 | 41.0 | 8,356 | 166 | 50.3 | 19.9 | < 0.001 |
| Target sensitivity | 95 | ||||||
| LR | 92.4 | 14.5 | 12,041 | 171 | 70.4 | Reference | Reference |
| DNN | 94.6 | 19.9 | 11,293 | 175 | 64.5 | 8.4 | < 0.001 |
DNN, deep neural network; ACRN, advanced colorectal neoplasia; NNScope, number needed to colonoscope to detect one ACRN; LR, logistic regression.