Table IX.
Performance of X‐Net compared with previous approaches on the same testing dataset.
| Author | Identification rate [%] | Localization Error [mm], mean (standard deviation) | ||||||||
|---|---|---|---|---|---|---|---|---|---|---|
| All | C | T | L | S | All | C | T | L | S | |
| H. Chen et al. 27 | 84.2 | 91.8 | 76.4 | 88.1 | n/a | 8.8(13.0) | 5.1 (8.2) | 11 (17) | 8.2 (8.6) | n/a |
| Yang et al. 28 | 85.0 | 92.0 | 81.0 | 83.0 | n/a | 8.6 (7.8) | 5.6 (4.0) | 9.2 (7.9) | 11 (11) | n/a |
| Liao et al. 18 | 88.3 | 95.1 | 84.0 | 92.2 | n/a | 6.5 (8.6) | 4.5 (4.6) | 7.8 (10.2) | 5.6 (7.7) | n/a |
| Sekuboyina et al. 16 | 86.7 | 89.4 | 83.1 | 92.6 | n/a | 6.3 (4.0) | 6.1 (5.4) | 6.9 (5.5) | 5.7 (6.6) | n/a |
| Sekuboyina et al. 16 | 87.7 | 89.2 | 85.8 | 92.9 | n/a | 6.4 (4.2) | 5.8 (5.4) | 7.2 (5.7) | 5.6 (6.2) | n/a |
| Sekuboyina et al. 16 | 88.5 | 89.9 | 86.2 | 91.4 | n/a | 6.2 (4.1) | 5.9 (5.5) | 6.8 (5.9) | 5.8 (6.6) | n/a |
| McCouat et al. 21 | 85.5 | 90.6 | 79.8 | 92.0 | n/a | 5.6 (7.1) | 3.9 (5.3) | 6.6 (7.4) | 5.4 (8.7) | n/a |
| J. Chen et al. 29 | 88.0 | n/a | n/a | n/a | n/a | 7.1 (7.1) | n/a | n/a | n/a | n/a |
| Jakubicek et al. 22 | 90.9 | n/a | n/a | n/a | n/a | 5.1 (4.0) | 4.21 (0.6) | 5.3 (1.3) | 6.6 (0.6) | n/a |
| Qin et al. 20 | 89.0 | 90.8 | 86.7 | 89.7 | 96.9 | 2.9 (5.8) | 2.2 (5.6) | 3.4 (6.5) | 2.9 (4.3) | 2.2 (2.7) |
| Y. Chen et al 30 | 94.7 | 89.5 | 95.3 | 100 | n/a | 2.6 (3.2) | 2.5 (3.7) | 2.6 (3.3) | 2.2 (1.8) | n/a |
| Proposed | 86.8 | 94.0 | 80.1 | 91.1 | 90.6 | 3.8 (2.9) | 3.3 (2.3) | 3.9 (3.0) | 3.7 (3.2) | 5.8 (3.7) |
| Yang et al. 28 * | 90.0 | 93.0 | 88.0 | 90.0 | n/a | 6.4 (5.9) | 5.2 (4.4) | 6.7 (6.2) | 7.1 (7.3) | n/a |
| Btrfly NetTL * | 87.1 | 86.6 | 86.5 | 90.4 | 84.4 | 4.1 (2.8) | 3.7 (2.4) | 4.5 (3.0) | 3.8 (2.7) | 4.9 (2.8) |
| ProposedTL * | 91.3 | 93.6 | 90.6 | 92.5 | 84.4 | 3.3 (2.7) | 3.0 (2.0) | 3.5 (2.8) | 2.8 (2.4) | 5.9(3.8) |
All, All regions; C, Cervical, T, Thoracic; L, Lumbar; S, Sacral. *Represents approaches which use data in addition to that provided by the MICCAI 2014 dataset. ProposedTL = X‐Net with transfer learning on a cohort of 803 patients; Yang et al. 28 *indicates the addition of patients outside the MICCAI 2014 dataset. All approaches listed in the table are 3D‐based except Btrfly Net 16 and X‐Net which are projection‐based. Scores from Btrfly Net without the addition of a general adversarial network are featured in the topmost row of approaches by Sekuboyina et al. The highest scoring metrics across all models (*with and without additional data) are in bold.