Fig 1.
Overview of proposed TACICP algorithm for carotid image registration.
The segmented 2D contours from images are the only inputs for our algorithm. Centerline and surface features are generated automatically from contours for two steps. The final output of the registration is a transformation composed by the rigid transformation Trigid from rigid initialization step and the thin-plate-spline transformation TTPS from non-rigid refinement step.
Fig 2.
Example of (a) match stage and (b) conditional match stage with the same point set pair at the same position.
The same color indicates the same category of points. The solid line represents the fixed point set, and the dash line represents the moving point set. The arrows are partial matching between two point sets.
Fig 3.
Example slices near the carotid bifurcation from a healthy volunteer for MERGE (left), SNAP (middle), and US (right) images.
Fig 4.
Comparison of different feature-based algorithms using average LMSD (a) and LMAXD (b) on US-MERGE and US-SNAP datasets.
An asterisk indicates statistically significant (p < 0.05) reduction in average LMSD or LMAXD as to TACICP algorithm.
Table 1.
RSR1.5 and computation time with the same configuration for different feature-based algorithms on US-MERGE and US-SNAP datasets.
Fig 5.
Lumen contours of CCA (orange), ICA (green) and ECA (red) in single MR (MERGE) slice with ICA and ECA (a) or CCA (b) using different feature-based algorithms.
The solid lines represent the MR contours, and the dash lines represent the transformed US contours drew on the same slice.
Fig 6.
Comparison of correct label configuration (dark) and reverted configuration (light) with average LMSD on US-MERGE and US-SNAP datasets.
An asterisk indicates statistically significant (p < 0.05) increment in average LMSD from correct configuration to reverted one.
Fig 7.
Comparison of ICP (dark) and CICP (light) in two steps with (a) average LMSD and (b) LMAXD on US-MERGE and US-SNAP datasets.
An asterisk in (a) and (b) indicates statistically significant (p < 0.05) reduction in average LMSD or LMAXD from ICP to CICP.
Table 2.
RSR1.5 of ICP and CICP algorithms in two steps on US-MERGE and US-SNAP datasets.
Fig 8.
Comparison of CICP with different steps using (a) average LMSD and (b) LMAXD on US-MERGE and US-SNAP datasets.
An asterisk in (a) and (b) indicates statistically significant (p < 0.05) reduction in average LMSD or LMAXD as to TACICP algorithm.
Fig 9.
Checkerboard views of single slice with ICA and ECA (a) or CCA (b) for MR (MERGE) and US images using CICP algorithm with different step combinations.
The patches in top left are from US images. US patches and MR patches appear alternately in the checkerboards. The arrows show the boundary between MR carotid lumen and US carotid lumen from CCA (orange), ICA (green) and ECA (red).
Fig 10.
Average LMSD with different conditional distance weight of CCA-ECA or CCA-ICA pair on US-MERGE (dark) and US-SNAP (light) datasets.
The dash line indicated the average LMSD with approximately infinite weight (w1 = 10000) on both datasets.
Fig 11.
Average ΔLMSD using TACICP algorithm with different amplitude of zero mean Gaussian noise on the contours on US-MERGE and US-SNAP datasets.
Fig 12.
Average LMSD using TACICP algorithm and rigid CICP algorithm for different relative distance from MR bifurcation slices (the closest slice to vessel bifurcation) on US-MERGE and US-SNAP datasets.
Fig 13.
Comparison of registration results of TACICP algorithm with the state-of-the-art intensity-based and hybrid algorithms using average LMSD on US-MERGE and US-SNAP datasets.
HYBRID represents hybrid model method. MI represents mutual information method.
Table 3.
Results for multi-contrast and multi-temporal registration using TACICP algorithm.
The second column represent the registration errors evaluated with average LMSD. The last column represents LMAXD.