Nonrigid registration of cardiac perfusion MR images using adaptive local template matching
View Patent ↗Motion of the heart during contrast enhancement is modeled at two hierarchical levels: global translation and local deformation. Large-scale, i.e., global, translational motion is identified using global template matching, while the local deformation of a portion of the object (e.g., the left ventricle where the object is a heart) is obtained using an adaptive local template matching scheme which adapts to the local edge distributions by varying the height and the width of the template as well as the location of the template followed by spatial regularization. To address the difficulty of rapid intensity changes across the image sequence as the contrast agent perfuse into the tissue, an edge-based similarity measure is used for template matching.
1. A method for registering a sequence of images of an object, comprising:
using medical imaging equipment to perform the steps of:
identifying motion of the object using global template matching;
obtaining local deformation of the object comprising:
defining a threshold G th for a gradient map and an area of a threshold area, A th for a local template;
obtaining a set of pixels, P, whose gradient magnitudes are greater than G th ;
for each pixel p(x, y) within a region of interest, selecting from P its K nearest neighbors N(x, y);
finding the minimum bounding box that encloses pixel p(x, y) as well as N(x, y);
setting the center of the local template as the center of the bonding box;
computing the area of the bounding box A N (x, y);
setting the area of the local template as A(x, y)=max(A N (x, y),A th );
setting the ratio between height and width of a template;
determining the height and the width of the local template for pixel (x,y); and;
estimating displacement of pixel (x, y) based on its local template.
2. The process recited in claim 1 wherein:
r
(
x
,
y
)
=
∑
(
x
,
y
)
∈
N
(
x
,
y
)
M
(
x
,
y
)
sin
(
θ
(
x
,
y
,
)
)
∑
(
x
,
y
)
∈
N
(
x
,
y
)
M
(
x
,
y
)
cos
(
θ
(
x
,
y
,
)
)
where: θ(x, y) and M(x, y) stand for respectively the orientation and the magnitude of the gradient.
3. The method recited in claim 2 including: applying weighted spatial regularization to the obtained local deformation.
4. The method recited in claim 3 the applying weighted spatial regularization to the obtained local deformation comprises:
obtaining a certainty map of the displacement estimation;
selecting a 2-dimensional Gaussian applicability function;
applying normalized convolution using the obtained certainty map and the Gaussian applicability function to obtain a regularized displacement field.
5. The method recited in claim 4 wherein the certainty map is C(dx,dy,t c )=max(|v(dx,dy,t r )| 2 ·S*(dx,dy,t c ), 0)
where S*(dx,dy,t c ) is the local maximum similarity measure in the current frame and |v(dx,dy,t r )| 2 is the gradient magnitude in the reference frame.
6. A method for registering a sequence of images of an object, comprising:
using medical imaging equipment to perform the steps of:
identifying motion of the object using global template matching;
obtaining local deformation of the object comprising:
estimating displacement of each pixel p(x, y) based on a local template having a variable size rectangular region centered at p(x,y) large enough to include a number of image gradient values above a predetermined threshold;
wherein the rectangular regions has dimensions in accordance with the image gradients within the region.
7. The method recited in claim 6 wherein rectangular region has an area A(x,y), and wherein the height and width of the rectangular region relative to a template r(x, y) is such that the height is the square root of (A(x,y)r(x,y)) and the width is the square root of (A(x,y)/r(x,y)).