Automated segmentation of a motion mask to preserve sliding motion in deformable registration of thoracic CT
Overview of the relevant anatomy of the thorax. The dashed line defines the contour of the motion mask.
Overview of the proposed method for extracting the motion mask. The figure shows sagittal views through the right lung of the results obtained for patient 1. The top row shows the input CT image, and a 3D surface rendering of the corresponding motion mask obtained using the method. The second row shows the label images, obtained by extracting anatomical features from the CT image. This yields the bony anatomy, the patient body and the lungs. The label images are then combined (+) and used to constrain (−) the evolving interface during consecutive level set processing steps, the results of which are shown in the bottom row. The current mask (white) is shown in overlay with the edges of the extracted features (black). From left to right are shown: the centered ellipsoid used to initialize the level set, the contour after reaching the detection point just in front of the anterior patient-to-air interface (the detection point—located in the center sagittal plane—projected onto this plane, would be located in the lower right corner of the image), the contour after having covered 95% of the lungs, and the final motion mask.
The final motion mask: (a) a coronal view of the final mask shown on the edges of the used mirrored version of the binary label images; (b) two axial views of the mask: the top one taken halfway through the lungs and the bottom one taken from the most inferior plane of the image; (c) 3D surface renderings of the anterior (top) and posterior (bottom) view of the motion mask.
Sagittal views of overlays of the exhale and inhale image pairs for three patients, corresponding to patients 1, 8, and 14. The contours of the motion mask extracted for each of the images is also shown for the exhale (inner) and the inhale image (outer).
Comparison of the results obtained for patient 1 for conventional registration, registration using a lung mask, and using the motion mask. Column (a) shows the difference of the reference and the target images when compensated with the obtained motion estimate, while (b) shows an enlarged view of the highlighted region, and (c) is the deformation vector field for that same region.
The mean target registration error with respect to the landmarks is plotted in function of the B-spline control point spacing. The error bars correspond to one standard deviation.
The distance between all landmarks in the lungs before registration (BR), the TRE after conventional registration without using a mask (NM), after registration using a lung mask (LM) and using the motion mask (MM). Given are the mean values (μ) and standard deviation (σ), and the maximum value over all points (Max).
The top part of the table shows the group mean of the distance between landmarks before registration (BR), of the TRE after conventional registration without using a mask (NM), after registration using a lung mask (LM) and using the motion mask (MM). The TRE is calculated for all points in the lung based on 3620 measurements (A), for all points within 10 mm of the chest wall based on 757 measurements (B), and for all points within 10 mm of the diaphragm and mediastinum using 636 landmarks (C). The bottom part of the table lists the DSC for the extracted bony anatomy (D), and the trachea and bronchi (E). Given are the mean values (μ) and standard deviation (σ).
Article metrics loading...
Full text loading...