US 11,386,591 B2 · Reconstruction method for motion compensated high quality 4D-CBCT image based on bilateral filtering

Abstract

The present disclosure provides a reconstruction method for motion compensated high quality four-dimensional (4D)-Cone Beam Computed Tomography (CBCT) image based on bilateral filtering, including the following steps: step 1, building a motion compensated three-dimensional (3D)-CBCT reconstruction model based on a simultaneous algebraic reconstruction technique (SART) to reconstruct a high quality CBCT image of a reference phase (phase 0%); step 2, creating an iterative optimization procedure and estimating a 4D Deformable Vector Field (DVF) model between phase 0% and other 4D phase images, thereby obtaining an exact 4D-DVF inclusive of an inverse sliding motion on a surface of a locomotive organ; and step 3, successively deforming the high quality phase 0% image in accordance with an optimized 4D-DVF to obtain a sequence of final high quality 4D-CBCT images. This method permits exact reconstruction of high quality 4D-CBCT images without changing the hardware structure of an existing linear accelerator for conventional radiotherapy.

A reconstruction method for motion compensated high quality four-dimensional (4D)-Cone Beam Computed Tomography (CBCT) image based on bilateral filtering, the method allowing for exact modeling reconstruction for an inverse sliding motion on a surface of a locomotive organ and comprising the following steps:

building a motion compensated three-dimensional (3D)-CBCT reconstruction model based on a simultaneous algebraic reconstruction technique (SART) to reconstruct a high quality CBCT image of a reference phase, the reference phase being phase 0%;
creating an iterative optimization procedure and estimating a 4D Deformable Vector Field (DVF) model between phase 0% and each other 4D phase image, thereby obtaining an exact 4D-DVF inclusive of the inverse sliding motion on the surface of the locomotive organ;
and
successively deforming the high quality phase 0% image in accordance with an optimized 4D-DVF to obtain a sequence of final high quality 4D-CBCT images;
wherein
DVF at each 4D phase is optimized by an algorithm that generates the optimized DVF by registration between 1) measured projections of the phase and 2) simulated forward projection of a reference phase 0% image deformed to the phase;
an energy function used in DVF optimization by projection registration is as follows:
f1(v0→t)=∥pt−Aμ0(x+v0→t)∥l22+βφ(v0→t)
f2(vt→0)=∥p0−Aμt(x+vt→0)∥l22+βφ(vt→0)
s.t.v0→t(x+vt→0)+vt→0=vt→0(x+v0→t)+v0→t=0
wherein f1 and f2 represent symmetric inverse continuous energy functions, with the superscript 0 representing phase 0% and the superscript t representing other phase t;
A is a matrix for providing a forward projection of the deformed phase 0% image;
v0→t represents a forward DVF;
vt→0 represents an inverse DVF;
∥pt−Aμ0(x+v0→t)∥l22 and ∥p0−Aμt(x+vt→0)∥l22 are data fidelity terms;
φ(v0→t) and φ(vt→0) are corresponding normalization terms;
the inverse continuity of DVF is guaranteed by the third row in the above equations;
the smoothness of the smoothing term is controlled by β;
and taking modeling of inverse sliding motions on surfaces of locomotive organs into consideration, φ(v) is defined as:
φ
⁡
(
v
)
=
∑
v
∈
R
3
⁢
⁢
∑
i
=
1
3
⁢
∑
j
=
1
,
y
j
∈
N
⁡
(
x
j
)
3
⁢
(
G
x
⁡
(
x
j
,
y
j
,
σ
x
2
)
·
G
μ
⁡
(
μ
t
⁡
(
x
j
)
,
μ
t
⁡
(
y
j
)
,
σ
μ
2
)
·
G
v
i
⁡
(
v
i
⁡
(
x
j
)
,
v
i
⁡
(
y
j
)
,
σ
v
2
)
⁢
(
∂
v
i
∂
x
j
)
2
)
where
∂
v
i
∂
x
j
represents the difference of a neighborhood of each voxel in DVF, j represents a latitude index of the voxel in the image, and i represents an index of the DVF along the three directions of x, y, and z;
specific forms of three sub-filter kernels are as follows:
⁢
G
x
⁡
(
x
j
,
y
j
,
σ
x
2
)
=
exp
⁡
(
-
(
x
j
-
y
j
)
T
⁢
(
x
j
-
y
j
)
2
⁢
σ
x
2
)
⁢
G
μ
⁡
(
μ
t
⁡
(
x
j
)
,
μ
t
⁡
(
y
j
)
,
σ
μ
2
)
=
exp
⁡
(
-

μ
t
⁡
(
x
j
)
-
μ
t
⁡
(
y
j
)

2
2
⁢
σ
μ
2
)
G
v
i
⁡
(
v
i
⁡
(
x
j
)
,
v
i
⁡
(
y
j
)
,
σ
v
2
)
=
exp
⁡
(
-
(
v
i
⁡
(
x
j
)
-
v
i
⁡
(
y
j
)
)
T
⁢
(
v
i
⁡
(
x
j
)
-
v
i
⁡
(
y
j
)
)
2
⁢
σ
v
2
)
Gx is the Gaussian kernel on the spatial domain with the variance of σx2—;
Gμ is another image domain based Gaussian kernel with the variance of σμ2;
Gv is the DVF domain Gaussian kernel with the variance of σv2;
‘xj’ is the kernel's center voxel location, and ‘yj’ is the neighborhood voxel locations within
N
⁡
(
x
j
)
;
∂
v
i
∂
x
j
represents a difference of a neighborhood of each voxel in DVF;
the gradient ∇φ(v)|v is calculated within the 3×3×3 neighborhood surrounding each voxel of interest;
the gradient of φ(v) is given by:
∇
φ
⁡
(
v
)
=
∑
v
∈
R
3
⁢
⁢
∑
i
=
1
3
⁢
∑
j
=
1
,
y
j
∈
N
⁡
(
x
j
)
3
⁢
(
G
x
⁡
(
x
j
,
y
j
,
σ
x
2
)
·
G
μ
⁡
(
μ
t
⁡
(
x
j
)
,
μ
t
⁡
(
y
j
)
,
σ
μ
2
)
·
G
v
⁡
(
v
i
⁡
(
x
j
)
,
v
i
⁡
(
y
j
)
,
σ
v
2
)
⁢
(
v
i
⁡
(
x
j
)
-
v
i
⁡
(
y
j
)
)
)
.

446 more elements

Reconstruction method for motion compensated high quality 4D-CBCT image based on bilateral filtering

Life of this patent

Reconstruction method for motion compensated high quality 4D-CBCT image based on bilateral filtering

Life of this patent

Abstract

Claims3