Table 2.
Overview of the Model Performance of the Convolutional Neural Network When the Independent Test Set (n = 479 with 180 Healthy Tooth Surfaces, 216 Noncavitated Carious Lesions, and 83 Cavitations) Was Used for Detection of Cavitations.
| True Positives | True Negatives | False Positives | False Negatives | Diagnostic Performance | ||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Detection of Cavitation | n | % | n | % | n | % | n | % | ACC | SE | SP | NPV | PPV | AUC |
| Results from all the included teeth and surfaces (n = 479 test images) | ||||||||||||||
| 25% of the images | 382 | 79.7 | 53 | 11.1 | 14 | 2.9 | 30 | 6.3 | 90.8 | 92.7 | 79.1 | 96.5 | 63.9 | 0.916 |
| 50% of the images | 381 | 79.6 | 61 | 12.7 | 15 | 3.1 | 22 | 4.6 | 92.3 | 94.5 | 80.3 | 96.2 | 73.5 | 0.931 |
| 75% of the images | 382 | 79.8 | 61 | 12.7 | 14 | 2.9 | 22 | 4.6 | 92.5 | 94.6 | 81.3 | 96.5 | 73.5 | 0.948 |
| 100% of the images | 381 | 79.5 | 66 | 13.8 | 15 | 3.1 | 17 | 3.6 | 93.3 | 95.7 | 81.5 | 96.2 | 79.5 | 0.955 |
| Results from anterior surfaces—incisors/canines (n = 153 test images) | ||||||||||||||
| 25% of the images | 106 | 69.3 | 23 | 15.0 | 7 | 4.6 | 17 | 11.1 | 84.3 | 86.2 | 76.7 | 93.8 | 57.5 | 0.887 |
| 50% of the images | 108 | 70.6 | 29 | 18.9 | 5 | 3.3 | 11 | 7.2 | 89.5 | 90.8 | 85.3 | 95.6 | 72.5 | 0.916 |
| 75% of the images | 109 | 71.2 | 27 | 17.7 | 4 | 2.6 | 13 | 8.5 | 88.9 | 89.3 | 87.1 | 96.5 | 67.5 | 0.932 |
| 100% of the images | 109 | 71.2 | 33 | 21.6 | 4 | 2.6 | 7 | 4.6 | 92.8 | 94.0 | 89.2 | 96.5 | 82.5 | 0.951 |
| Results from posterior surfaces—molars/premolars (n = 326 test images) | ||||||||||||||
| 25% of the images | 276 | 84.7 | 30 | 9.2 | 7 | 2.1 | 13 | 4.0 | 93.9 | 95.5 | 81.1 | 97.5 | 69.8 | 0.941 |
| 50% of the images | 273 | 83.7 | 32 | 9.8 | 10 | 3.1 | 11 | 3.4 | 93.6 | 96.1 | 76.2 | 96.5 | 74.4 | 0.932 |
| 75% of the images | 273 | 83.7 | 34 | 10.4 | 10 | 3.1 | 9 | 2.8 | 94.2 | 96.8 | 77.3 | 96.5 | 79.1 | 0.967 |
| 100% of the images | 272 | 83.4 | 33 | 10.1 | 11 | 3.4 | 10 | 3.1 | 93.6 | 96.5 | 75.0 | 96.1 | 76.7 | 0.957 |
| Results from vestibular and oral surfaces—anterior/posterior teeth (n = 225 test images) | ||||||||||||||
| 25% of the images | 155 | 68.9 | 39 | 17.3 | 12 | 5.3 | 19 | 8.5 | 86.2 | 89.1 | 76.5 | 92.8 | 67.2 | 0.884 |
| 50% of the images | 156 | 69.3 | 43 | 19.1 | 11 | 4.9 | 15 | 6.7 | 88.4 | 91.2 | 79.6 | 93.4 | 74.1 | 0.923 |
| 75% of the images | 160 | 71.1 | 41 | 18.2 | 7 | 3.1 | 17 | 7.6 | 89.3 | 90.4 | 85.4 | 95.8 | 70.7 | 0.937 |
| 100% of the images | 159 | 70.7 | 47 | 20.9 | 8 | 3.5 | 11 | 4.9 | 91.6 | 93.5 | 85.5 | 95.2 | 81.0 | 0.943 |
| Results from occlusal surfaces—molars/premolars (n = 253 test images) | ||||||||||||||
| 25% of the images | 227 | 89.7 | 13 | 5.1 | 2 | 0.8 | 11 | 4.4 | 94.9 | 95.4 | 86.7 | 99.1 | 54.2 | 0.939 |
| 50% of the images | 225 | 88.9 | 17 | 6.7 | 4 | 1.6 | 7 | 2.8 | 95.7 | 97.0 | 81.0 | 98.3 | 70.8 | 0.914 |
| 75% of the images | 222 | 87.7 | 19 | 7.5 | 7 | 2.8 | 5 | 2.0 | 95.3 | 97.8 | 73.1 | 96.9 | 79.2 | 0.962 |
| 100% of the images | 222 | 87.7 | 18 | 7.1 | 7 | 2.8 | 6 | 2.4 | 94.9 | 97.4 | 72.0 | 96.9 | 75.0 | 0.966 |
The calculations were performed for different types of teeth, surfaces, and training steps, which resulted in different subsamples.
ACC, accuracy; AUC, area under the receiver operating characteristic curve; SE, sensitivity; SP, specificity; NPV, negative predictive value; PPV, positive predictive value.