Table 2.
Average 3D Euclidean distance (mm) by landmarks and by registration methods with the first observer as anatomical reference.
| Landmark Distance Error (mm) | |||||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Bones | Soft tissues | Precision (AVG) | Accuracy (AVG SD) | ||||||||||
| 1 | 2 | 3 | 4, 5 | 6 | 7 | 8 | 9 | 10 | 11, 12 | 13, 14 | |||
| Interobserver variability | 1.00 | 1.24 | 1.40 | 2.21 | 2.43 | 2.56 | 1.27 | 1.36 | 1.85 | 2.19 | 3.15 | 2.01 | 1.29 |
| FFD MI filtered CTs | 1.18 | 2.60 | 1.45 | 1.62 | 3.03 | 2.89 | 2.64 | 2.07 | 2.11 | 3.85 | 2.64 | 2.44 | 1.30 |
| Demons MI filtered CTs | 1.13 | 2.40 | 1.72 | 1.77 | 2.93 | 2.97 | 2.27 | 2.84 | 2.89 | 3.63 | 2.79 | 2.54 | 1.33 |
| Demons MI | 1.13 | 2.36 | 1.80 | 1.82 | 3.03 | 3.12 | 2.36 | 2.79 | 2.87 | 3.79 | 2.82 | 2.59 | 1.38 |
| Demons MSE | 1.15 | 2.79 | 1.81 | 1.65 | 3.42 | 3.21 | 3.02 | 2.72 | 2.58 | 3.82 | 3.84 | 2.81 | 1.63 |
| FFD MSE filtered CTs | 1.32 | 3.49 | 1.62 | 1.91 | 3.50 | 3.27 | 3.63 | 2.94 | 2.19 | 4.06 | 3.10 | 2.86 | 1.54 |
| FFD MI | 1.07 | 3.81 | 1.93 | 1.77 | 3.37 | 3.52 | 3.96 | 3.11 | 2.07 | 4.16 | 3.07 | 2.92 | 1.60 |
| Demons MSE filtered CTs | 1.35 | 3.05 | 1.80 | 1.82 | 3.60 | 3.63 | 3.13 | 2.78 | 2.92 | 3.80 | 4.25 | 3.00 | 1.64 |
| FFD MSE | 1.83 | 5.13 | 3.30 | 2.50 | 4.07 | 4.37 | 5.14 | 5.05 | 2.63 | 5.42 | 3.48 | 3.88 | 2.18 |
| Demons D. maps | 2.43 | 5.51 | 4.62 | 2.79 | 4.40 | 4.75 | 5.17 | 5.97 | 4.01 | 6.35 | 3.41 | 4.43 | 2.14 |
| FFD D. maps | 2.40 | 6.03 | 4.02 | 3.15 | 4.42 | 4.80 | 5.94 | 6.37 | 4.37 | 6.68 | 3.55 | 4.65 | 2.33 |
| Rigid MSE | 3.16 | 6.26 | 4.32 | 4.16 | 4.85 | 5.11 | 6.26 | 6.57 | 4.69 | 6.85 | 4.53 | 5.16 | 2.52 |
AVG: average, AVG SD: average standard deviation, FFD: free form deformation, MSE: mean squared error, MI: mutual information, and CT: computed tomography (scan).
Methods are classified by their performance order. The performance is defined by the accuracy (average Euclidean distance error) and the precision (average of the standard deviation Euclidean distance error). A second observer allows quantifying the interobserver variability.
The “FFD with MI on filtered CTs” errors are inferior to all the methods errors (P < 0.05), except for the “demons with MI on filtered CTs” and “demons with MI on original CTs” methods (P > 0.15). The “demons with MI on filtered CTs” errors are inferior to all the methods errors (P ≤ 0.03) except for the “demons with MI on original CTs,” and the “FFD on filtered CTs” for both MSE and MI metrics (resp., P = 0.24, = 0.08 and = 0.76).