Table 4.
Mean ± standard deviation results comparing two-dimensional (2D) U-Net and U-Net++ segmentations on knee osteoarthritis patient 2D transverse ultrasound images and subsequent three-dimensional (3D) reconstructed segmentations to the corresponding manual gold standard segmentations. The p-value corresponds to a comparison between 2D and 3D segmentation performance.
| Method | Segmentation | Precision (%) | Recall (%) | DSC (%) | MSD (mm) | HD (mm) | A/VPD (%) |
|---|---|---|---|---|---|---|---|
| U-Net | 2D Transverse | 83.9 ± 2.6 | 94.7 ± 1.7 | 88.9 ± 1.6 | 0.23 ± 0.03 | 1.28 ± 0.35 | 13.4 ± 9.4 |
| 3D Reconstruction | 73.2 ± 8.6 | 94.3 ± 1.3 | 82.3 ± 6.1 | 0.44 ± 0.13 | 1.91 ± 0.57 | 30.1 ± 15.6 | |
| P-value | 0.001 | 0.598 | 0.007 | <0.001 | 0.099 | 0.014 | |
| U-Net++ | 2D Transverse | 84.5 ± 5.8 | 93.0 ± 1.0 | 88.4 ± 3.5 | 0.24 ± 0.03 | 1.18 ± 0.43 | 10.4 ± 7.6 |
| 3D Reconstruction | 72.7 ± 8.0 | 95.0 ± 1.1 | 82.2 ± 5.7 | 0.43 ± 0.11 | 2.02 ± 0.64 | 31.4 ± 13.9 | |
| P-value | 0.002 | 0.005 | 0.011 | 0.037 | 0.038 | 0.007 |