Table 1b:
Segmentation results for each structure averaged over the 5 outlier test patients. We provide the standard deviation and range of the results. The best performance for each metric is in bold.
| Model | Metric | GTV | G. maximus | G. medius | G. minimus | Paraspinal |
|---|---|---|---|---|---|---|
| 3D U-Net Dice | VDSC | 81.1 ± 7.4 (69.5,90.7) | 85.4 ± 6.6 (73.8,93.3) | 88.1 ± 4.6 (79.0,91.6) | 83.7 ± 3.4 (77.9,86.7) | 78.9 ± 10.4 (67.9,91.8) |
| 3D U-Net Cross-entropy | 81.4 ± 6.0 (73.0,90.9) | 87.2 ± 2.8 (83.2,91.7) | 88.3 ± 3.9 (80.7,91.1) | 81.2 ± 3.7 (74.4,84.4) | 79.5 ± 9.6 (69.5,92.2) | |
| 3D U-Net Weighted cross-entropy | 82.7 ± 6.8 (72.8,93.0) | 85.0 ± 5.2 (77.0,91.1) | 87.8 ± 4.1 (80.6,91.7) | 82.5 ± 3.2 (76.7,86.1) | 78.0 ± 15.4 (50.5,92.8) | |
| Residual 3D U-Net Dice | 82.0 ± 6.9 (72.6,90.9) | 87.1 ± 3.9 (81.1,93.0) | 88.9 ± 3.4 (82.6,92.2) | 81.9 ± 4.2 (74.5,86.1) | 84.7 ± 7.2 (71.9,93.2) | |
| Residual 3D U-Net Cross-entropy | 78.4 ± 6.8 (68.2,88.3) | 85.4 ± 2.6 (81.9,88.4) | 84.8 ± 4.8 (75.5,88.7) | 79.2 ± 6.5 (66.7,84.8) | 81.9 ± 9.8 (64.0,92.1) | |
| Residual 3D U-Net Weighted Cross-entropy | 78.8 ± 5.0 (71.4,86.2) | 86.5 ± 3.6 (80.2,91.3) | 88.5 ± 3.1 (83.2,92.0) | 81.7 ± 4.4 (74.2,87.4) | 85.7 ± 7.0 (74.7,92.8) | |
| Average Ensemble | 82.1 ± 6.4 (72.9,91.7) | 87.9 ± 3.1 (82.4,91.5) | 89.4 ± 3.3 (83.1,92.0) | 83.6 ± 3.6 (77.2,87.5) | 83.8 ± 7.0 (75.0,93.1) | |
| Optimal Ensemble | 81.9 ± 6.6 (72.6,91.8) | 87.8 ± 3.1 (82.4,91.1) | 89.4 ± 3.2 (83.3,92.1) | 83.4 ± 3.9 (76.4,87.7) | 84.9 ± 6.6 (75.0,93.2) | |
| 3D U-Net Dice | SDSC 2 mm | 41.7 ± 9.6 (27.1,52.6) | 71.8 ± 11.7 (54.6,89.4) | 82.8 ± 5.6 (72.0,87.5) | 84.8 ± 4.3 (77.2,90.3) | 68.2 ± 17.2 (40.0,87.6) |
| 3D U-Net Cross-entropy | 40.8 ± 8.5 (29.6,52.8) | 73.8 ± 7.9 (66.0,87.5) | 83.0 ± 5.1 (73.3,87.2) | 80.7 ± 6.0 (70.1,87.5) | 68.5 ± 16.6 (45.9,90.2) | |
| 3D U-Net Weighted cross-entropy | 46.9 ± 10.9 (32.8,64.3) | 71.2 ± 8.4 (60.4,84.7) | 81.6 ± 5.9 (72.3,89.6) | 82.7 ± 5.6 (75.6,90.6) | 69.2 ± 22.4 (33.6,92.1) | |
| Residual 3D U-Net Dice | 45.3 ± 10.2 (34.3,61.7) | 76.2 ± 8.6 (66.6,88.4) | 83.9 ± 6.7 (72.1,91.7) | 83.4 ± 6.1 (74.2,90.3) | 78.5 ± 11.6 (61.4,93.3) | |
| Residual 3D U-Net Cross-entropy | 35.5 ± 7.3 (26.4,45.5) | 72.0 ± 7.1 (64.5,82.1) | 74.3 ± 7.8 (59.8,82.0) | 77.7 ± 9.6 (60.5,89.2) | 74.3 ± 13.4 (50.7,89.9) | |
| Residual 3D U-Net Weighted Cross-entropy | 43.8 ± 6.4 (35.1,52.4) | 76.8 ± 6.6 (69.1,88.5) | 81.6 ± 6.1 (71.2,87.9) | 81.9 ± 6.0 (72.3,87.8) | 81.2 ± 9.6 (64.2,92.0) | |
| Average Ensemble | 44.0 ± 8.7 (31.0,55.9) | 76.7 ± 7.7 (67.2,88.8) | 85.0 ± 5.3 (75.1,89.7) | 85.0 ± 5.5 (76.1,91.0) | 76.5 ± 13.4 (56.6,92.4) | |
| Optimal Ensemble | 44.2 ± 8.7 (31.6,56.4) | 76.9 ± 7.4 (67.9,88.4) | 84.9 ± 5.5 (74.7,90.1) | 84.9 ± 5.8 (75.4,90.9) | 79.3 ± 10.6 (66.3,93.0) | |
| 3D U-Net Dice | SDSC 3 mm | 49.2 ± 10.8 (33.3,62.8) | 78.3 ± 11.0 (61.5,94.1) | 88.5 ± 5.1 (78.3,92.2) | 90.4 ± 3.3 (84.7,94.1) | 74.6 ± 15.9 (49.4,92.5) |
| 3D U-Net Cross-entropy | 48.1 ± 9.6 (35.9,62.6) | 80.1 ± 6.9 (72.4,91.9) | 88.6 ± 4.4 (80.3,92.4) | 87.1 ± 5.1 (78.6,92.2) | 74.9 ± 16.0 (55.5,94.5) | |
| 3D U-Net Weighted cross-entropy | 54.0 ± 11.7 (39.2,73.3) | 76.7 ± 9.3 (63.4,90.3) | 87.4 ± 5.2 (79.1,94.0) | 88.4 ± 4.8 (83.1,94.9) | 74.3 ± 22.0 (38.1,96.0) | |
| Residual 3D U-Net Dice | 52.4 ± 11.6 (39.3,69.6) | 81.8 ± 7.8 (71.5,93.0) | 89.2 ± 5.8 (79.6,95.6) | 89.4 ± 5.3 (81.9,94.3) | 83.5 ± 10.8 (64.9,96.3) | |
| Residual 3D U-Net Cross-entropy | 42.1 ± 8.3 (31.6,51.9) | 78.0 ± 5.8 (70.9,86.2) | 81.6 ± 7.1 (68.8,88.2) | 84.3 ± 8.9 (67.9,93.9) | 79.7 ± 13.7 (54.2,93.7) | |
| Residual 3D U-Net Weighted Cross-entropy | 50.0 ± 7.4 (39.8,60.5) | 82.3 ± 6.1 (73.6,92.4) | 87.7 ± 5.2 (78.6,92.5) | 87.4 ± 5.5 (79.4,92.4) | 85.8 ± 9.7 (68.2,95.6) | |
| Average Ensemble | 51.3 ± 10.0 (36.8,66.1) | 82.7 ± 6.8 (73.1,92.7) | 90.2 ± 4.5 (82.0,94.4) | 90.5 ± 4.5 (84.0,94.7) | 82.1 ± 11.7 (67.4,95.9) | |
| Optimal Ensemble | 51.4 ± 10.1 (37.1,66.6) | 82.8 ± 6.7 (73.2,92.4) | 90.1 ± 4.7 (81.7,94.6) | 90.3 ± 4.9 (83.2,94.7) | 84.6 ± 9.6 (69.9,96.3) | |
| 3D U-Net Dice | ASSD | 7.1 ± 3.5 (3.5,13.0) | 2.8 ± 1.4 (1.0,5.0) | 1.9 ± 1.2 (1.1,4.2) | 1.3 ± 0.2 (1.0,1.7) | 2.5 ± 1.3 (1.1,4.4) |
| 3D U-Net Cross-entropy | 6.8 ± 3.1 (3.3,12.1) | 2.8 ± 1.2 (1.3,4.9) | 1.8 ± 0.8 (1.2,3.3) | 1.6 ± 0.4 (1.2,2.2) | 4.2 ± 3.9 (1.0,11.4) | |
| 3D U-Net Weighted cross-entropy | 6.3 ± 3.1 (2.5,11.2) | 6.2 ± 7.1 (1.4,20.3) | 2.3 ± 1.1 (1.1,4.3) | 1.4 ± 0.3 (1.0,1.8) | 3.0 ± 2.7 (0.9,8.1) | |
| Residual 3D U-Net Dice | 6.8 ± 3.3 (3.1,12.0) | 2.4 ± 0.9 (1.1,3.8) | 1.7 ± 0.8 (1.0,3.1) | 1.5 ± 0.4 (1.0,2.2) | 2.8 ± 3.0 (0.8,8.7) | |
| Residual 3D U-Net Cross-entropy | 8.1 ± 3.6 (4.3,14.4) | 3.3 ± 1.5 (1.9,6.3) | 2.4 ± 0.8 (1.7,3.9) | 1.9 ± 0.8 (1.1,3.3) | 3.2 ± 3.3 (1.0,9.8) | |
| Residual 3D U-Net Weighted Cross-entropy | 7.6 ± 2.8 (4.6,12.2) | 2.5 ± 0.9 (1.4,4.1) | 1.7 ± 0.5 (1.2,2.6) | 1.5 ± 0.4 (1.2,2.2) | 1.8 ± 1.1 (0.9,3.9) | |
| Average Ensemble | 6.9 ± 3.4 (3.0,12.5) | 2.3 ± 0.8 (1.4,3.5) | 1.6 ± 0.7 (1.0,2.9) | 1.3 ± 0.3 (0.9,1.8) | 2.0 ± 1.1 (0.8,3.9) | |
| Optimal Ensemble | 7.0 ± 3.4 (3.0,12.6) | 2.3 ± 0.8 (1.4,3.6) | 1.6 ± 0.7 (1.0,2.9) | 1.3 ± 0.3 (0.9,1.8) | 1.9 ± 1.1 (0.8,3.9) |