Introduction
-
We present a novel comparative visualization approach based on interactive derivation of locally rigid transforms in a region of interest from a pre-computed whole-body deformable registration. These are subsequently applied to the coordinated visualization of baseline, follow-up, and fused whole-body slices. As bones are inherently rigid, these transforms form a good basis for compensating for posture differences by matching follow-up to baseline by simply selecting one bone at a time. Using this locally rigid transform approach, baseline and follow-up anatomical regions can be interactively matched, based on a single click of the mouse in the anatomical region of interest; see Fig. 1 for an illustration.×
-
Besides visualizing the synchronized side-by-side baseline and matched follow-up slices, we present four ways to facilitate a details-on-demand-based visual comparison of the two datasets: 1) The deformation sphere [13] is a spherical cloud of points that follows the mouse cursor, with each of the points being warped by the local deformation field. The sphere assists comprehension of the actual local deformation field. 2) The color fusion view combines the baseline and follow-up in one view by making use of two (perceptually motivated) color channels. 3) We render a set of uncertainty iso-contours around the current region of interest [14], explicitly visualizing the approximation error of the locally rigid transform. 4) We have integrated a magic lens [15] that displays an alternate MR modality so that areas of change can be further studied.
-
Finally, we investigate the role and utility of our methods toward improving the radiologists’ workflow in the assessment of tumor burden in patients suffering from Kahler’s disease by means of a case study with two experienced skeletal radiologists.
Related work
Integrated exploration through locally rigid transforms
Locally rigid registration
-
It is based on rigid image matching: no distortions are introduced.
-
It requires minimal user interaction: one mouse click is sufficient for aligning the regions of interest.
-
It does not require manual segmentation of the regions: the latter is calculated automatically, starting from the user-provided seed point (one mouse click). Subsequently, the corresponding region of the follow-up scan is obtained from the pre-computed global deformation.
-
It is fast: the more time-consuming operations are performed at the pre-processing stage. Segmentation of the structure of interest, estimation of the rigid transform in the area, and re-alignment of the follow-up volume are performed on-the-fly. Derivation of the locally rigid transform from the global deformation field is done with a closed-form relation, which will always be significantly faster than any iterative on-the-fly locally rigid registration technique.
Computing localized transforms
Area of interest segmentation
Baseline and follow-up point matching
Rigid estimation by landmark transform
Application of found rigid transform
User interface
STIR viewing
Color fusion view
Deformation sphere
Uncertainty contours overlay
Implementation and performance
Pre-processing
Stitching and inhomogeneity correction
Registration
elastix
[37]. Using this software library, we have constructed a multi-level registration approach. Initially, the coarse alignment is estimated using an affine transform. After this initial registration, an elastic registration takes place. This elastic registration makes use of a multi-resolution process (3 resolutions), in which a deformable B-spline grid is incrementally refined. In the last pass, the finest level, the B-spline grid has a spacing of 12 mm. As a similarity metric to be optimized by the registration, we have chosen to use a Mattes mutual information [19] (32 histogram bins), although other metrics may also be suitable for this purpose. A stochastic gradient descent optimizer [38] was selected, using 1,000 iterations.Timing and performance
Operation | Time |
---|---|
Preprocessing (bias correction and stitching) | 10–15 min |
Preprocessing (registration) |
\(<\)5 min |
Sphere deformation | 10 ms |
Segmentation | 100–200 ms |
Slicing | 100–200 ms |
Evaluation
Case study
Interactive locally rigid transforms enable the rapid matching of slices from baseline and follow-up.
The deformation sphere helps to clearly distinguish between areas of rigid and non-rigid change.
The deformation sphere helps to gain insight into the nature of localized non-rigid changes, more so than standard side-by-side comparison for example.
The confidence boundaries clearly show where the rigid approximation is justified, and hence where the visual differences can be trusted.
The presented system speeds up whole-body MRI comparison both through the rapid matching of slices and through the subsequent explicit visual representation of local changes (deformation sphere/color fusion).
The STIR magic lens aids lesion assessment, in that lesion details can be better visualized with STIR, whilst (detailed) context is shown by the \(T_1W\).
The color fusion view facilitates visual comparison between baseline and follow-up more than side-by-side views, as changes are visually emphasized.