DIR for contour propagation
Pre-treatment and repeat 4D respiratory correlated PET/CT scans together with a 3D intra-venous contrast-enhanced CT scan were obtained for 17 NSCLC patients. The study was approved by the Institutional Review Board. The repeat 4DCT scan was obtained during the second week of treatment, typically around the 8
th day after the start of radiation therapy, the details of which can be found in Van Elmpt et al. [
19]. The voxel sizes for all CT scans used in this study were 0.98 mm × 0.98 mm in plane, with 3 mm slice thickness. The 50% exhale phase image was used for structure delineation in both scans. The primary tumor (GTV), lungs, esophagus and spinal cord were delineated by an expert physician on each of the pre- and mid-treatment scans. The heart was typically not contoured in the clinical workflow for these patients so was not included in this study. The nodal-GTV was delineated on 12/17 patients using FDG-PET image data for determination of lymph node involvement. The remaining 5/17 patients did not have nodal-GTVs. Pre-treatment delineation of involved lymph nodes was then performed using a contrast-enhanced CT image. Mid-treatment delineation was performed using the mid-ventilation phase (50% exhale) of a 4DCT image. DIR was performed on the mid-ventilation phase of the 4DCT, deforming the pre-treatment to the mid-treatment scans and applying the resulting deformation map to the pre-treatment scan ROIs to obtain ROIs on the mid-treatment scan. Prior to deformation, rigid registration was performed using an automated local correlation algorithm to improve the initial registration used as input to the DIR algorithms.
Three DIR algorithms were used – Demons (Fast Symmetric) [
20], Salient Feature Based Registration (SFBR) [
21], as implemented in a research version of the Pinnacle™ RTPs (v9.100, Philips Medical Systems, Fitchburg, WI), and Morphons [
22]. The Demons algorithm was a modified version of that used in the Insight Segmentation and Registration Toolkit (ITK) [
23]. The Demons algorithm uses a regular grid of forces to deform an image to a target image based on matching intensity values between two images. The optical flow equation is used to derive the displacement orientation and magnitude. The Fast Symmetric algorithm used in this study is a multi-resolution approach whereby a maximum of 200, 100, 100 and 30 iterations are run at each resolution level from 8× the image resolution to 1× the image resolution respectively. At each resolution level the standard deviation of the Gaussian smoothing was 3, 3, 0.9 and 0.7 mm for resolution levels 8 times to 1 times the image resolution. Histogram matching was performed with 64 levels and 7 match points.
The SFBR algorithm is an automated version of landmark-based registration in which sharply prominent and distinctive features are automatically located in each image. The features are found using an interest point detector algorithm and can be detected at any point in the patient. Typically 1000–2000 features are obtained. Once the features have been located, they are assigned a location determined by the position of their centroid as well as a scale. The salient feature locations are then used as anchor points to interpolate a non-rigid transformation using the Thin Plate Splines (TPS) method.
The Morphons algorithm uses quadrature phase differences to estimate deformations between two images [
22,
24]. Quadrature phase differences describe local structure such as edges between dark and bright areas in an image. Due to the use of quadrature phase differences this method is invariant to image intensity and weak gradients [
25]. A multi-resolution implementation of Morphons was used applying 8 resolution steps with the final resolution equal to the full resolution of the CT scan with at maximum 20 iterations per resolution step, and at maximum 4 for the final grid size. Gaussian smoothing was applied between all resolution steps with a standard deviation of 1.25 times the voxel size of the resolution grid. The Morphons algorithm was implemented in Matlab (R2009a, The Mathworks, Natick, Ma) [
25].
The Demons and Morphons algorithms result in a deformation vector field (DVF) at the image resolution, in the frame of reference of the second (during treatment) image. DIR was performed by deforming the pre-treatment scan to the mid-treatment scan. Propagation of the ROIs is then performed by looking up the value of the DVF in each voxel in second image and tracing it back to the pre-treatment image and obtaining the value of the binary mask of each ROI at that location. ROIs were propagated using the same technique for both the Demons and Morphons DVF. The SFBR algorithm results in a set of TPS equations that can be evaluated at any point in the image. For propagation, the source ROIs are converted to meshes and the TPS is evaluated at each vertex location in the meshes to determine the location of each mesh in the second image. No smoothing of the propagated ROIs was performed for any algorithm.
The DIR-propagated ROIs were compared with the physician-drawn ROIs on the mid-treatment scan using the Dice score and the mean slicewise Hausdorff distance (MSHD). The Dice score for two ROIs A and B was defined as 2|A∩B|/(|A|+|B|). The MSHD is the average over all slices of the largest value of the smallest distance to agreement between two ROIs on each slice [
26]. The difference in the Centre of Mass (COM) position of the GTVs was also measured, as this has implications on isocenter placement in adaptive re-planning. A one-way Analysis of Variance (ANOVA) test was carried out to determine the statistical significance of any differences between the two algorithms. A value of
p = 0.05 was used as the threshold for statistical significance.
In addition to the quantitative metric scores, the ROIs were evaluated qualitatively by an expert physician to determine the clinical utility of the algorithms for ROI propagation. A score of 1 was given to ROIs that were clinically acceptable without modification, 2 was given to ROIs that were clinically useful but required minor modification on several slices, and 3 was given to ROIs that were not clinically useful, where it would be more efficient to start the contouring from scratch.