Table 2.
Performance metrics of the keypoint region-based convolutional neural network (KR-CNN) including external and internal validation for automated measurement of thoracic Cobb angles in the validation cohort of patients undergoing long-segment posterior spinal fusion for adolescent idiopathic scoliosis correction
| Model | MAE ± SD (95% CI) | MedAE (IQR; 95% CI) | MSE | SMAPE | ICC GT |
|---|---|---|---|---|---|
| KR-CNN Resnet 50 (Current Study, External Validation) | 2.22 ± 2.04 (1.06 to 3.39) | 1.47 (0.73–2.97; 0.89 to 3.15) | 9.1 | 4.29 | 0.98 |
| KR-CNN Resnet 50 (Current Study, Internal Validation) | 1.78 ± 1.61 (1.12 to 2.44) | 1.21 (0.65–2.43; 0.81 to 2.74) | 6.84 | 3.67 | 0.99 |
| Landmark Net | 13.85 | — | — | 38.87 | — |
| Angle Net | 9.36 | — | — | 23.91 | — |
| AEC-Net | 4.90 | — | — | 23.59 | — |
| Vertebra Landmark Detection | — | — | — | 10.81 | — |
| Auto-CA | — | — | — | 5.27 | — |
| VLTENet | 2.51 | — | — | 5.44 | — |
| RegNet | 2.92 | — | — | 6.87 | |
| MVE-Net | 7.81 | 24.94 | |||
| Seg4Reg | — | — | — | 21.71 | — |
| ScolioVis | — | — | — | 8.97 | — |
| Minimum bounding rectangle | — | — | — | — | 0.97 |
ICC Intraclass correlation coefficient, GT ground truth, IQR interquartile range, MAE Mean absolute error, MedAE Median absolute error, MSE Mean squared error, KR-CNN Keypoint Region-based convolutional neural network, Resnet 50 residual network 50, SD Standard deviation, SMAPE Symmetric mean absolute percentage error