Table 3.
A comparison of detection and segmentation performance on RibFrac Test Set, of FracNet, two deep neural network counterparts (3D FCN and 3D DeepLab), two radiologists (R1 and R2) and their union.
| Methods | Detection Sensitivities @ FP Levels |
Detection |
Segmentation |
|||||||
|---|---|---|---|---|---|---|---|---|---|---|
| 0.5 | 1 | 2 | 4 | 8 | Avg | Sensitivity | Avg FP | IoU | Dice | |
| FracNet | 66.0% | 75.0% | 81.7% | 90.5% | 92.9% | 81.2% | 92.9% | 5.27 | 55.6% | 71.5% |
| 3D FCN | 59.9% | 69.7% | 76.1% | 84.4% | 87.8% | 75.6% | 87.8% | 7.02 | 49.1% | 66.2% |
| 3D DeepLab | 63.7% | 72.5% | 79.2% | 88.2% | 91.3% | 79.0% | 91.3% | 6.11 | 50.3% | 68.7% |
| R1 | / | / | / | / | / | / | 79.1% | 1.34 | 47.4% | 64.3% |
| R2 | / | / | / | / | / | / | 75.9% | 0.92 | 36.7% | 53.1% |
| R1 R2 | / | / | / | / | / | / | 83.1% | 1.80 | 47.8% | 64.7% |
| FracNet R1 | / | / | 83.9% | 90.4% | 93.8% | 82.6% | 93.8% | 5.99 | 54.9% | 70.9% |
| FracNet R2 | / | / | 85.8% | 92.6% | 95.7% | 84.4% | 95.7% | 5.83 | 52.5% | 68.9% |