1
Consistency-based Respiratory Motion Estimation in Rotational
2
Angiography
3
Mathias Unberath∗ and Andreas Maier
4
Pattern Recognition Lab, Computer Science Department,
5
Friedrich-Alexander-University Erlangen-Nuremberg and
6
Erlangen Graduate School in Advanced Optical Technologies (SAOT), Germany
7
Andr´e Aichert
8
Pattern Recognition Lab, Computer Science Department,
9
Friedrich-Alexander-University Erlangen-Nuremberg
10
Stephan Achenbach
11
Department of Cardiology, Friedrich-Alexander-University Erlangen-Nuremberg
12
(Dated: March 11, 2017)
13
Purpose: Rotational coronary angiography enables 3D reconstruction but suffers
14
from intra-scan cardiac and respiratory motion. While gating handles cardiac mo-
15
tion, respiratory motion requires compensation. State-of-the-art algorithms rely on
16
3D-2D registration that depends on initial reconstructions of sufficient quality. We
17
propose a compensation method that is applied directly in projection domain. It
18
overcomes the need for reconstruction and thus complements the state-of-the-art.
19
Methods: Virtual single-frame background subtraction based on vessel segmenta-
20
tion and spectral deconvolution yields non-truncated images of the contrasted lu-
21
men. This allows motion compensation based on data consistency conditions. We
22
compensate craniocaudal shifts by optimizing epipolar consistency to (a) devise an
2 23
image-based surrogate for cardiac motion and (b) compensate for respiratory mo-
24
tion. We validate our approach in two numerical phantom studies and three clinical
25
cases.
26
Results: Correlation of the image-based surrogate for cardiac motion with the
27
ECG-based ground truth was excellent yielding a Pearson correlation of 0.93 ± 0.04.
28
Considering motion compensation, the target error measure decreased by 98 % and
29
69 %, respectively, for the phantom experiments while for the clinical cases the same
30
figure of merit improved by 46 ± 21 %.
31
Conclusions: The proposed method is entirely image-based and accurately esti-
32
mates craniocaudal shifts due to respiration and cardiac contraction. Future work
33
will investigate experimental trajectories and possibilities for simplification of the
34
single-frame subtraction pipeline.
∗
E-mail:
[email protected]; web: www5.cs.fau.de/~unberath
3 I.
35
INTRODUCTION
36
Minimally invasive treatment of cardiovascular disease (CVD) heavily relies on inter-
37
ventional X-ray imaging [1]. Traditionally, coronary angiograms are acquired from few
38
viewing angles only, impeding assessment due to projective simplifications, such as overlap
39
and foreshortening [2–4]. These limitations can be alleviated by providing physicians with
40
3D reconstructions of the arterial tree, which is expected to improve diagnostic assessment
41
and interventional guidance [1, 5, 6]. Unfortunately, the low temporal resolution of C-arm
42
cone-beam CT (CBCT) scanners inhibits straight-forward 3D reconstruction. Rotational
43
sequences are acquired over several seconds and, therefore, are corrupted by cardiac and
44
respiratory motion. Consequently, both motion patterns have to be incorporated into re-
45
construction algorithms.
46
Cardiac motion is highly complex but repetitive with high frequency. Instead of estimating
47
a full 4D motion field, retrospective gating of the sequence is sufficient. After gating, only
48
images of the same cardiac phase contribute to the reconstruction. Gating approaches are
49
generic in the sense that they can be employed in both back-projection-type [7, 8] and sym-
50
bolic reconstruction algorithms [3, 4, 9]. Moreover, the choice of low-dimensional surrogate
51
signal is unrestricted. Simultaneously acquired electrocardiograms (ECG) are common sur-
52
rogate signals [8, 10]. Yet, image-based surrogates exist [3] and should be preferred as they
53
encode the motion-state rather than the electro-physiologic activation.
54
In contrast to cardiac motion, respiratory motion is low frequency such that only very few
55
recurrences are observed during a standard acquisition.
56
ning strategies followed by cyclic registration have proved robust in image-guided radiation
57
therapy [11–15]. Unfortunately, in the context of rotational angiography respiratory phase
Respiratory phase gating or bin-
4 58
binning is impossible as it would result in a highly ill-posed problem, impeding the appli-
59
cation of methods proposed by Sonke et al. [11], Li et al. [12], and Brehm et al. [13, 14]. A
60
prominent solution is to avoid the problem as a whole by requiring patients to hold their
61
breath throughout the acquisition. If breath-hold is impossible or imperfect, motion estima-
62
tion techniques must be applied to mitigate the corruption due to respiration and, finally,
63
enable 3D reconstruction.
64
State-of-the-art methods for respiratory motion compensation in standard protocols require
65
an initial 3D reconstruction of the structure of interest, that enables 3D/2D registration of
66
the model to subsequent motion states. In bi-plane angiography it has been shown that
67
this strategy allows for the estimation of 3D rotations and translations [16] but even dense
68
3D displacement fields can be recovered [17]. Single-plane rotational acquisitions also offer
69
multiple views enabling reconstruction. Blondel et al. used alternating reconstruction and
70
bundle adjustment steps to estimate detector shifts to account for respiratory motion [3, 18].
71
In the context of rotational angiography, however, motion estimation is far more challeng-
72
ing as supposedly consistent frames are acquired successively rather than simultaneously.
73
Hence, there is no guarantee that a particular motion state is observed more than once
74
which would prevent any uncompensated reconstruction in the first place. Aforementioned
75
approaches are similar in that they seek to establish correspondences among different views
76
by transformation of a 3D model. In case of non-residual, low frequency motion, the recon-
77
struction of such initial model is challenging or even impossible, making its requirement a
78
limiting factor for the applicability of the methods.
79
The good news is that the motion of the heart due to respiration is considerably less com-
80
plex than cardiac motion itself. It is linearly related to diaphragm motion [19] and can
81
be approximated as a 3D translation that exhibits its largest component in craniocaudal
5 82
direction [3, 20, 21]. In practice, the craniocaudal axis of the patient aligns with the vertical
83
detector coordinate such that the optimization for detector domain shifts should already
84
yield substantial improvement [3]. To omit the need for initial reconstructions, we build on
85
a recently published motion estimation technique that is applied in projection domain and
86
optimizes for consistency among all images. The method exploits redundancies in transmis-
87
sion imaging to derive data consistency conditions (DCC) from the epipolar geometry [22].
88
Unfortunately, the method cannot be applied to truncated images as truncation inherently
89
voids the consistency of an acquisition. Although the projection of the whole thorax is trun-
90
cated, the coronary arteries do not extend beyond the field-of-view. Separation of the con-
91
trasted coronary arteries from the remaining thorax yields non-truncated projections that
92
could be used as input to DCC-based motion estimation algorithms. A straight-forward
93
method to achieve material decomposition is background removal via digital subtraction.
94
For this task, projections of the same scene without contrast enhancement are estimated
95
from the acquired images and a segmentation mask [23]. The virtual background image
96
is subtracted from the original projections yielding non-truncated images of the contrast
97
agent only. Background suppression is an integral component of many vessel reconstruction
98
methods but is usually low-level and used for artifact reduction only [6]. Extending its use-
99
fulness to purely projection-based motion detection and compensation is, therefore, of high
100
relevance.
101
In this paper we show that virtual subtraction imaging of the coronary arteries enables
102
consistency-based respiratory motion estimation. Epipolar consistency is used to estimate
103
vertical detector shifts in numerical cone-beam CT simulations of an XCAT-like phantom
104
[24, 25] and on real patient data. Finally, the high frequency component of the estimated
105
shifts that is correlated to the superior-inferior motion of the heart during contraction [21]
6 106
is used to derive an image-based surrogate for cardiac motion.
II.
107
METHODS
108
We describe an algorithm targeted at translational motion estimation in rotational cone-
109
beam CT angiography using epipolar consistency conditions (ECC). First, the contrasted
110
vessels must be separated from the background via virtual subtraction imaging in order to
111
obtain non-truncated images and, therefore, enable consistency-based processing. Second,
112
horizontal and vertical shifts of the detector are estimated that maximize consistency among
113
all images in the sense of epipolar consistency conditions. Finally, we extract high frequency
114
components of the estimated motion patterns that correlate with the superior-inferior motion
115
of the heart during contraction to derive an image-based surrogate for the cardiac phase.
116
A.
Virtual subtraction imaging
117
Background removal is already used in many vessel reconstruction algorithms to reduce
118
artifact or to enable sparsity priors. To this end, simple image processing techniques such
119
as morphological closure [3] or top-hat filtering are sufficient [8]. For consistency-based pro-
120
cessing we seek to reliably separate the contrasted vessels from the background via digital
121
subtraction. Therefore, high fidelity estimates of the contrasted regions are needed. Subse-
122
quently, the segmented pixels are removed from the images and then filled with values that
123
are estimated from a local neighborhood. The estimated background is then subtracted
124
from the original projection, ideally yielding a non-truncated image of the contrasted vessels
125
only. A schematic of the algorithm is given in Figure 1.
7
Subtraction Segmentation algorithm
Inpainting algorithm
Masking
X-ray projection
Virtual mask image
Binary mask
Digital subtraction angiogram
FIG. 1. Schematic of the truncation removal. A binary segmentation of the object is used to remove the corresponding pixels and the background is estimated using an inpainting method. Finally, the background is subtracted from the original projection yielding the difference image.
1.
126
Segmentation
127
In X-ray imaging, contrasted vessels appear as narrow tubular structures that are bright
128
with respect to the surrounding background. It seems natural to exploit this prior knowledge
129
for segmentation in a multi-scale approach based on first and second order derivatives. The
130
acquired images I 0 are preprocessed using the well-known bilateral filter that suppresses
131
noise while preserving edges [26].
132
Let α(u, σ) be the local orientation of a structure at position u ∈ R2 and scale σ, and let eα
133
be a unit vector orthogonal to the respective orientation. If the structure at u is a contrasted
134
vessel of scale σ we expect high negative gradients in the direction orthogonal to α as well
135
as a high negative curvature along eα . Filters that enhance structures with aforementioned
136
properties are already known in literature. We briefly restate vessel enhancement and Koller
137
filtering following [27] and [28], respectively. Finally, we describe a strategy to merge both
138
responses in a binary segmentation mask.
139
In order to determine the respective responses at position u, first the Hessian matrix is
8 140
computed as
H(u, σ) =
141
∂ 2 Iσ (u) ∂u21
∂ 2 Iσ (u) ∂u1 ∂u2
∂ 2 Iσ (u) ∂u2 ∂u1
∂ 2 Iσ (u) ∂u22
,
(1)
142
where Iσ = I ∗Gσ , Gσ is a 2D Gaussian kernel with standard deviation σ, I is the projection
143
image after bilateral filtering, and ∗ denotes convolution. The Eigenvalues of the Hessian
144
matrix λ1/2 , |λ1 | ≥ |λ2 | are related to the tubularity of an underlying structure and allow for
145
the definition of meaningful quantities. The Blobness Rb = λ2/λ1 is close to zero for tubular
146
147
structures and close to unity for bright regions with similar extent in all directions, such p as blobs. The structureness S = λ21 + λ22 is given by the Frobenius norm of the Hessian
148
and is close to zero for uniform background and large in regions with high contrast [27].
149
Combining the two measures into a probability-like response yields the vesselness
V(u) = max exp(−c1 Rb ) · (1 − exp(−c2 S)) ,
150
σ
(2)
where c1/2 are constants that control the influence of Rb and S. Tuning c1/2 for sensitivity yields strong responses for vessels, but also for other structures with high curvature. Subsequent hysteresis thresholding and removal of small connected components suppresses isolated responses and yields a binary mask Vb . To remove some of the non-vessel segmentations still included in the binary mask the Koller filter can be used. We found it to have a higher specificity than the vesselness, yet, Koller filtering only yields strong responses on the centerline of arteries. Given the Hessian H(u, σ), we can compute the orientation α(u, σ) of the generating structure as tan(2α) = 2H12 (H11 − H22 ). For vessels with scale σ we expect large negative gradients at its boundaries u ± σeα , yielding the Koller filter response n n > > oo K(u) = max min − ∇Iσ (u + σeα ) eα , ∇Iσ (u − σeα ) eα . σ
9
TABLE I. Heuristically determined parameters used in the segmentation algorithms. Constant c2 is defined using the c-th percentile of the structureness Pc (S)). As the Koller filter response is not normalized, tK is dependent on image domain gradients and, hence, the contrast level. c1
c2
c3
tK
Ph1 and Ph2
1/0.52
1/(P0.94 (S))2
4.0 mm
0.5
R1, R2,and R3}
1/0.52
1/(P0.98 (S))2
4.0 mm
0.07
151
Thresholding of K yields an approximation of the set of centerline points QK = {u | K(u) > tK },
152
where tK is a heuristically determined threshold. We define g(u) = min{ku − vk2 } with
153
v ∈ QK as the distance of position u to its closest centerline point, and merge both responses
154
in a binary segmentation mask W such that (
155
W(u) =
1 for Vb (u) 6= 0 ∧ g(u) < c3
,
(3)
0 otherwise 156
where c3 is chosen to reflect the largest vessel radius to be expected.
157
The performance of aforementioned algorithms depends on the choice of parameters that
158
were tuned heuristically. Table I summarizes the values used for phantom and real data
160 159
experiments. Finally, we define the inverse segmentation mask W, where W(u) = 1−W(u).
161
2.
Background estimation
162
The segmentation masks obtained according to Section II A 1 are used to remove intensity
163
information at the locations of contrasted vessels, thereby artificially corrupting the projec-
164
tions. Formally, the corrupted image G is obtained by multiplying the original image with
165
the inverse mask such that G = I · W. The acquired image after noise filtering I and the
166
background image that we seek to estimate B are similar in that their corrupted versions
10 167
cannot be distinguished: !
168
G =I ·W =B·W .
169
The artificially corrupted observation G does not contain any information about the presence
170
of contrast agent. Consequently, the estimation solely relies on the remaining structures and
171
the resulting background estimate can be considered a virtual non-contrast image [23]. The
172
corrupted regions of B are narrow and sparse but they are connected. Simple methods, such
173
as morphological closure, may be rewarding for defect pixel interpolation [29], however, in
174
the case of background recovery they often yield patchy images with unnatural appearance.
175
A concurrent approach proposed by Aach et al. tries to recover the missing values using
176
spectral domain techniques [30]. According to the convolution theorem, Equation 4 can be
177
restated as a convolution in frequency domain g =
178
in the background image, and ∗ denotes convolution. Both G and B are real valued such that
179
their Fourier transforms g and b , respectively, are symmetric. Therefore, b (k) = b ∗ (N − k)
180
for a spectral line pair b (k) and b (N −k). Let the spectrum b of the undistorted background
181
B consist of only a single spectral line pair b (k) = bˆ(s) δ(k − s) + bˆ(N − s) δ(k − (N − s)).
182
Then, the observed line pair at s and N − s reads:
g (s) =
1 N
1 N
(4)
1 ( N
b ∗ w ), where N is the number of pixels
bˆ (s) · w (0) + bˆ ∗ (s) · w (2s)
(5)
183
g (N − s) = g ∗ (s) =
bˆ ∗ (s) · w ∗ (0) + bˆ (s) · w ∗ (2s)
184
where bˆ(s) and bˆ(N − s) = bˆ∗ (s) are the coefficients to be recovered. Equation 5 can be
185
solved for bˆ(s) yielding
186
bˆ(s) = N ·
g (s)w (0) − g ∗ (s)w (2s) . |w (0)|2 − |w (2s)|2
(6)
11 187
In practice, however, b consists of more than a single spectral line pair. After one such
188
deconvolution step the error is zero only at s and N − s. Therefore, deconvolution is applied
189
iteratively, estimating multiple line pairs s(i) and N − s(i) in succession. The line pairs
190
for each iteration are selected such that they maximize the decrease in Mean-Square Error
191
[30]. We apply spectral deconvolution in a patch wise manner to preserve the locality of
192
image appearance. The patches are weighted with a Blackman window to create artificially
193
continuous signals and to avoid spectral leakage [31].
194
Finally, the estimated background image is subtracted from the acquired images yielding
195
an artificial digital subtraction angiogram (aDSA) D(u) = I(u) − B(u) [32].
196
perfect segmentation and background estimation, D contains contributions of the contrast
197
agent only. As the contrast agent is confined to the coronary arteries that do not extend
198
beyond the field of view, the resulting aDSA images D are, ideally, not truncated.
199
enables the application of consistency-based methods for intra-scan motion detection that
200
are described in the following sections.
201
B.
Assuming
This
Epipolar consistency
202
Consistency-based processing requires a series of non-truncated images Di , i = 1, . . . , M
203
acquired in a known geometry that is defined by its projection matrices Pi ∈ R3×4 [22, 33].
204
The epipolar geometry describes the geometric relation between two views i1 and i2 , where
205
i1/2 ∈ {1, . . . , M } and i1 6= i2 . It allows for the definition of corresponding lines lii21 and
206
lii12 in the images Di1 and Di2 , respectively. The lines arise from the intersection of the
207
respective epipolar plane E, which is a homogeneous 3-vector, with the two image planes.
208
A schematic of the described relations is given in Figure 2. The resulting lines in the domain
209
of the respective images are defined by a signed distance to the image center t and an angle
12 210
β measured with respect to the u1 axis, such that l = (− sin(β), cos(β), −t)> .
211
Assuming an orthogonal acquisition geometry, integration over corresponding epipolar lines
212
lii21 and lii12 gives two redundant ways of calculating the integral over the epipolar plane E
213
through the three dimensional object. It holds
214
ρi1 lii21 − ρi2 lii12 = 0 ,
215
where ρi (li ) denotes integration over the line li in image Di . Equation 7 also provides
216
intuition of why subtraction imaging as discussed in Section II A is necessary: In case of
217
truncation the integrals over epipolar lines would not yield the complete plane integral
218
and would, therefore, disagree even when no motion is present. For cone-beam projections
219
aforementioned equality does not apply as integrals over epipolar lines differ by a weighting
220
with the distance to the camera center [22]. Yet, they carry redundant information. Aichert
221
et al. use Grangeat’s theorem to cancel out the weighting and derive consistency conditions
222
similar to Equation 7 for cone-beam geometries:
(7)
223
∂ ∂ ρi1 lii21 − ρi2 lii12 ≈ 0. ∂t ∂t
224
In Equation 8, t is the distance of the line to the image origin. There exists a pencil of
225
epipolar planes Eii21 (κ) around the baseline, where κ is the angle of the epipolar plane and
226
the principal ray and the baseline is the join of the two camera centers [22]. This allows for
227
the definition of a global consistency metric EC by summation over all image pairs:
228
EC =
M X X i1 =1 i2 fresp
Γ(f ) if
0
,
(11)
else
250
max is a constant related to the maximum where Γ denotes the Fourier transform of γ. fresp
251
respiratory and minimum cardiac frequency to be expected and set to 0.5 Hz.
252
max = 0.5 Hz proves reasonable considering an average ventilation rate of 0.33±0.08 Hz [34]. fresp
253
Moreover, the acquisition protocol discussed here is performed using breath-hold commands
254
suggesting residual respiratory motion.
255
correspond to the same cardiac phase and interpolate between them to obtain a normalized
256
cardiac time ci ∈ [0, 1] for each frame i. Then, we can calculate gating weights wi for a
257
reference heart phase cr following [8]
258
wi (cr ) =
a
cos
0
Setting
We detect peaks in γcard that are assumed to
|ci −cr | π w
if |ci − cr |
, k ∈ {i2 , i3 } maps to the blue epipolar lines. Images i1 and i2 correspond to the same cardiac phase cr . The rightmost image i3 , however, corresponds to a drastically different phase and shows residual mismatch between the compensated epipolar line and the tracked key point. This observation highlights the influence of cardiac motion and motivates the use of gating weights in Equation 14.
294
state the Structural Similarity (SSIM) and the Root-Mean-Square Error (RMSE) of the
295
background estimate with respect to the non-contrasted, noise-free ground-truth. We also
296
state SSIM and RMSE of noisy realizations of the non-contrasted ground-truth to establish
297
an upper bound for both measures. To avoid bias introduced by pixels that never changed,
298
the measures are computed only in a region of interest that is obtained by dilating the
299
estimated segmentation mask by 10 pixels.
300
C.
Motion compensation
301
Evaluating the performance of motion estimation is difficult. Direct comparison of the
302
estimated shifts with ground truth is unrewarding, as cardiac motion is non-rigid and, hence,
303
cannot be compensated for by the proposed strategy. To overcome this fundamental limi-
304
tation, we evaluate the compensation error for different target heart phases separately. To
18 305
this end, we track the 2D position ξij ∈ R2 of key points j = 1, . . . , Q, such as vessel bifur-
306
cations, over all frames i = 1, . . . , M . In the phantom experiments the reference positions
307
are obtained by forward projection of the key points in 3D. For the real data experiments
308
vessel bifurcations are tracked manually in 2D. Finally, the error averaged over all tracked
309
points at a given heart phase cr reads Q
M
310
X 1 1 X P wi (cr ) · ζij , εcr = Q j=1 i wi (cr ) i=1
311
where the gating weights wi (cr ) are calculated from γ according to Section II C. Moreover,
312
ζij is the weighted average distance of point j to the corresponding epipolar lines given by ζij
313
> X 1 j j k k > wk (ci ) · ξk · Fi · ξi + (0, ti , 0) , =P k6=i wk (ci ) k6=i
(13)
(14)
314
where we use underline notation to denote homogeneous extension, tki = γi − γk , and F ∈
315
R3×3 is the fundamental matrix that can be computed from the projection matrices [22].
316
The uncompensated error can be computed by setting tik = 0 in Equation 14. We also state
317
the Pearson correlation coefficient of the image-based surrogate for cardiac motion derived
318
according to Section II C with the gold standard surrogate that is derived from the ECG
319
signal.
IV.
320
A.
321
RESULTS
Segmentation and inpainting
322
Quantitative results of the segmentation quality assessment in the phantom study are
323
stated in Table IV. Sensitivity and specificity values are slightly smaller than unity indi-
324
cating minor segmentation errors. Representative examples of such mis-segmentations are
325
highlighted in Figure 4.
19
TABLE IV. Sensitivity and specificity for the segmentations of phantom data sets Ph1 and Ph2 at both photon fluence levels in the rectangular region of interest around the arteries with respect to the ground-truth. 5 · 102 photons
1 · 104 photons
Study Sensitivity
Specificity
Sensitivity
Specificity
Ph1
0.96 ± 0.017
0.99 ± 0.0033
0.96 ± 0.014
0.98 ± 0.0036
Ph2
0.96 ± 0.015
0.99 ± 0.0037
0.96 ± 0.014
0.99 ± 0.0036
326
SSIM and RMSE of the background estimation algorithm described in Section II A 2 can be
327
found in Table V and Table VI for the phantom studies with low and high photon counts per
328
pixel, respectively. Values of the SSIM and RMSE are worse for the more noisy realizations.
329
B.
330
Motion estimation
331
Ground truth motion patterns and the shifts estimated according to Section II B for
332
the phantom studies are shown in Figure 5. The cardiac motion is non-rigid which can be
333
understood from the dashed black lines in Figure 5 that correspond to the displacement
334
in u2 coordinate of the distinct bifurcation points ξ j of the artery tree. Qualitatively, the
335
estimated shifts are in very good agreement with the ground truth motion patterns.
336
For all phantom and real data sets, the high frequency component of the estimated shifts
337
correlated very well with cardiac motion. The Pearson correlation coefficients between sur-
338
rogates derived from the ECG and shifts γ are stated in Table VII. Correlation was higher
339
for the clinical cases.
340
Finally, we used the surrogates to assess the motion compensation error at distinct heart
R3
R2
R1
Ph2: 1·104 phot.
Ph1: 5·102 phot.
20
FIG. 4. From left to right: projection, segmentation, estimated background, and virtual subtraction image. Red and green arrows point to excess and missing segmentations, respectively. We show representative angiograms of the left coronary artery at different viewing angles for the two phantom data sets Ph1 and Ph2, and the three real data cases R1, R2, and R3. Individual data sets occupy one row each.
21
TABLE V. Quantitative results of the image inpainting step for phantom studies Ph1 and Ph2 with a photon fluence of 5 · 102 photons per pixel. We state the mean value and standard deviation of the structural similarity (SSIM) and the root-mean-squared error (RMSE) of the inpainted image Bˆ compared to the noise-free ground-truth. The values stated in brackets were obtained when comparing the a bilateral filtered version of the background estimate (see Section II A 1) to the noise-free ground-truth. Moreover, we provide an upper bound to aforementioned quantities by computing SSIM and RMSE for a noisy version of the ground-truth (noisy GT). @ @
SSIM
RMSE
inpainted
0.91 ± 0.03 (0.99 ± 0.01)
0.36 ± 0.04 (0.23 ± 0.06)
noisy GT
0.92 ± 0.03 (1.0 ± 0.0)
0.27 ± 0.03 (0.058 ± 0.006)
inpainted
0.91 ± 0.03 (0.99 ± 0.00)
0.36 ± 0.05 (0.23 ± 0.07)
noisy GT
0.92 ± 0.03 (1.0 ± 0.0)
0.27 ± 0.03 (0.057 ± 0.006)
@ @ Study @ @
Ph1
Ph2
341
phases according to Equation 13. The resulting errors are shown in Figure 7 and Figure 8
342
for the phantom and real data studies, respectively. We observed substantial reductions in
343
εcr for all studied cases, the numeric values are stated in Table VIII.
344
V.
345
346
A.
DISCUSSION
Segmentation and inpainting
347
From the phantom images shown in Figure 4 it becomes apparent that the proposed
348
pipeline yields segmentation masks that are marginally thicker than the underlying vessels.
22
TABLE VI. Inpainting results for both phantom studies when simulating 1 · 104 photons per pixel. Mean value and standard deviation of the SSIM and the RMSE of the inpainted image Bˆ and a bilateral filtered version (in brackets) compared to the noise-free ground-truth. Again, upper bounds for all measures are computed using a noisy version of the ground-truth (noisy GT). @ @
SSIM
RMSE
inpainted
0.98 ± 0.01 (0.99 ± 0.01)
0.26 ± 0.08 (0.25 ± 0.08)
noisy GT
1.0 ± 0.0 (1.0 ± 0.0)
0.060 ± 0.006 (0.034 ± 0.006)
inpainted
0.99 ± 0.01 (0.99 ± 0.00)
0.25 ± 0.09 (0.24 ± 0.09)
noisy GT
1.0 ± 0.0 (1.0 ± 0.0)
0.059 ± 0.006 (0.033 ± 0.005)
@ @ Study @ @
Ph1
Ph2
TABLE VII. Pearson correlation of the image-based surrogate for cardiac motion with the surrogate derived from the ECG. Study
Ph1
Ph2
Photons
5·102
1·104
5·102
1·104
Pearson R
0.88
0.88
0.91
0.91
R1
R2
R3
0.92
0.98
0.94
349
This effect arises from hysteresis thresholding and decreases specificity. Sensitivity is in-
350
fluenced by incomplete segmentations, e.g., at bifurcations. Although over-segmentation is
351
the dominant source of error, the impact of under-segmentation is more pronounced due
352
to the imbalanced foreground and background class labels.
353
slight decrease in specificity for phantom study Ph1when comparing high and low photon
354
count realizations, which suggests over-segmentation. This unexpected deterioration can be
355
attributed to the complex vessel enhancing filters, e.g., due to excess points in QK that have
356
responses barely higher than the allowed threshold tK . However, not having found sub-
Interestingly, we observed a
23 5·102 phot. γ + 10 mm 1·104 phot. γ + 5 mm Ground-truth
Shift in mm
30 20 10 0 -10 -20 1
1.5
2
2.5
3
3.5
4
4.5
5
5.5
Frame time in multiples of cardiac time
Shift in mm
(a) Study Ph1 5·102 phot. γ + 10 mm 1·104 phot. γ + 5 mm Ground-truth
10
0
1 1.5 2 2.5 3 3.5 4 4.5 5 5.5 6 6.5
Frame time in multiples of cardiac time
(b) Study Ph2
FIG. 5. Shifts γ obtained by optimizing for consistency are plotted for phantom studies Ph1 and Ph2 in the top and bottom figure, respectively. Shifts corresponding to different noise realizations are slightly offset to allow for improved visualization. The black, dotted curves show reference shifts of the vessel key points that were tracked in 3D.
357
stantial difference in performance for high and low noise realizations suggests that bilateral
358
filtering effectively mitigated the influences of noise (see Table IV).
359
more challenging in the clinical cases suggesting that the numerical phantom study was
360
limited with respect to anatomical variation and imaging physics. Obtaining good segmen-
361
tation quality was particularly difficult in views that exhibit overlap of the arterial tree with
362
the spine which drastically decreased vessel visibility. Moreover, we observed discontinuous
363
segmentations that arise from low vessel contrast.
Segmentation proved
Shift in mm, phase in a.u.
24
6
γ Cardiac phase
4 2 0 -2 -4 -6 20
40 60 80 100 Projection number
120 133
Shift in mm, phase in a.u.
(a) Clinical case R1 γ Cardiac phase
4 2 0 -2 -4 -6 20
40 60 80 100 Projection number
120 133
Shift in mm, phase in a.u.
(b) Clinical case R2 12 10 8 6 4 2 0 -2 -4 -6
γ Cardiac phase
20
40 60 80 100 Projection number
120 133
(c) Clinical case R3
FIG. 6. Shifts γ that maximize data consistency for the three clinical cases. Moreover, we show the corresponding cardiac time as derived from the ECG. Local maxima of the cardiac time coincide with R-peaks in the ECG and indicate ventricular systole.
364
Considering background estimation, we expect a deterioration of SSIM and RMSE for the
365
more noisy realizations. This is because the background estimate B is compared to noise-
366
free ground truth. The quantitative measures improved when noise-filtered [26] background
Error in mm
25 5·102 phot. εcr 1·104 phot. εcr No Comp. −10 mm
6 5 4 3 2 1 0.5 0 0
0.2
0.4
0.6
0.8
1
Reference heart phase
(a) Study Ph1 5·102 phot. εcr 1·104 phot. εcr No Comp.
Error in mm
2
1 0.5 0 0
0.2
0.4
0.6
0.8
1
Reference heart phase
(b) Study Ph2
FIG. 7. Compensation errors εcr at different reference heart phases cr . Note that the error curve for study Ph1 was shifted by −10 mm to improve visualization.
367
estimates were used indicating that the observed deteriorations originate, at least partly,
368
from noise characteristic. The proposed method was not able to achieve RMSE lower than
369
0.25 although theoretically possible (see right-most column in Table VI) indicating minor
370
inpainting errors. Yet, SSIM and RMSE of the background estimate are very close to the
371
respective upper bounds obtained using the noisy ground-truth. This observation suggests
372
that background estimation worked very well in most cases.
373
Segmentation errors propagate through the whole pipeline and erroneously segmented struc-
26 3
εcr No Comp.
Error in mm
2.5 2 1.5 1 0.5 0
0.2
0.4
0.6
0.8
1
Reference heart phase
(a) Clinical case R1
Error in mm
2.5
εcr No Comp.
2
1.5
1
0.5 0
0.2
0.4
0.6
0.8
1
Reference heart phase
(b) Clinical case R2
Error in mm
5
εcr No Comp. −2 mm
4
3 2.5 2 1.5 1 0
0.2
0.4
0.6
0.8
1
Reference heart phase
(c) Clinical case R3
FIG. 8. Compensation errors εcr at different reference heart phases cr for the clinical cases. Note that the error curve for study R3 was shifted by −2 mm to improve visualization.
27
TABLE VIII. Mean and minimal error with and without compensation stated in mm. For the phantom cases, we limit the analysis to the high photon count realization as the compensation errors are either the same or higher (see Figure 7). Ph1
Orig.
Ph2
Mean
Min.
Mean
Min.
14 ± 1
13
1.2 ± 0.6
0.39
Comp. 0.81 ± 0.23 0.26 0.75 ± 0.39 0.12
R1
R2
R3
Mean
Min.
Mean
Min.
Mean
Min.
1.9 ± 0.3
1.3
1.7 ± 0.2
1.4
6.2 ± 0.1
6.0
Comp. 1.0 ± 0.1 0.85 1.1 ± 0.1 0.95 2.7 ± 0.5
1.8
Orig.
374
tures, such as vertebrae, are pronounced in the binary mask as can be seen in the last row
375
of Figure 4. However, the presence of mis-segmentation is not as distinct in the subtraction
376
image, decreasing its influence on motion estimation. Moreover, we have found ECC to be
377
robust against statistical errors introduced in the subtraction pipeline which is supported
378
by the results presented in the following section.
379
B.
Motion estimation
380
From Figure 5 it is apparent that the respiratory motion modeled in study Ph1 is much
381
more drastic than for Ph2. A similar observation can be made when considering the clinical
382
cases shown in Figure 6, where cases R1 and R2 exhibit residual respiratory motion only
383
while scan R3 suffers from heavy intra-scan respiratory motion.
28 384
The surrogate for cardiac motion exhibited a higher correlation for the clinical cases. This
385
arises from the shape of the local maxima in γ that denote the start of ventricular con-
386
traction and are detected by the method detailed in Section II C: for the phantom data we
387
observed broader maxima that complicated exact localization (see Figure 5), whereas the
388
peaks signaling ventricular systole were sharp for the clinical cases, particularly for R1, as
389
can be seen in Figure 6. Small deviations in peak location heavily deteriorate Pearson corre-
390
lation as the surrogates have the form of sawtooth waves. Exemplarily, shifting a single peak
391
position to a neighboring frame yields a deterioration in Pearson correlation of 4.4 %. When
392
considering the motion compensation error at specific heart phases, we observed the lowest
393
error around ventricular diastole (cr ∼ 0.4 − 0.6). This result is to be expected, as cardiac
394
motion inside the gating window is negligible at diastole. However, some cases exhibit a
395
second local minimum at end systole that also coincides with an inflection point in γ. This
396
behavior is best observed in Figure 7b at cr ∼ 0.85 for the phantom study and in Figure 8b
397
at cr ∼ 0.9 for the clinical cases. In presence of drastic respiratory motion as in clinical case
398
R3, aforementioned behavior is suppressed. This observation is reflected in a high mean
399
error with a low standard deviation as can be understood from Table VIII and Figure 8c.
400
It is worth mentioning that the quantitative errors reported in Table VIII and Figures 7 and
401
8 depend on the choice of gating parameters (see Section II C), as they influence the gating
402
weights used in Equation 13. Wider gating windows increase the residual motion contained
403
in each gate which would, therefore, result in higher errors. The values a = 4 and w = 0.4
404
that were used here constitute a reasonable compromise between heart phase sensitivity and
405
residual cardiac motion [8, 35].
406
The reduction in error achieved for the clinical cases is substantial, yet an average minimal
407
error of 1.2 ± 0.5 mm (see Table VIII) at diastole remains. This residual error is a composite
29 408
of three different sources of error: incorrect estimates γi , a finite gating window width, but
409
also inaccuracies in manual key point tracking. Incorrect key point positions impose a fun-
410
damental lower bound on error reduction, unfortunately, direct assessment of this systematic
411
error is not possible for the real data cases. The results obtained on phantom data, however,
412
suggest that very accurate motion compensation even below the pixel size is possible with
413
the proposed algorithm.
C.
414
Limitations and challenges
415
First, the current optimization described in Section II B does not inhibit a systematic
416
shift of all detectors. Such shift is bounded by the maximally occurring motion and would
417
introduce cone-beam artifacts, the effect of which is negligibly compared to the drastic
418
motion artifacts. However, the method is, in principle, able to optimize for systematic drifts
419
[36], an extension that should be considered for future work.
420
Second, within this work we optimized for craniocaudal shifts only, a motion model that is
421
not sufficient in the most general case. However, respiratory motion of the coronary arteries
422
occurring in craniocaudal direction is substantially larger than in left-right or posterior-
423
anterior direction, respectively [21]. Compensation of the dominant motion component as
424
proposed here may allow for initial reconstructions and, therefore, enable methods based on
425
3D/2D registration that allow for more complex motion models [3, 16, 20].
426
Finally, the experiments were restricted to incomplete breathing cycles, an assumption that
427
was motivated by imperfect breath-hold. In previous work, we showed that the proposed
428
method is fit to handle multiple recurrences [23], yet, a demonstration on real data remains
429
subject of future work.
30 VI.
430
CONCLUSIONS
431
In summary, we proposed a projection domain algorithm targeted at intra-scan motion
432
assessment in rotational coronary angiography. The method is based on virtual single-
433
frame subtraction imaging, that uses segmentation and background estimation algorithms
434
to obtain non-truncated images of the contrast-filled lumen. Then, craniocaudal shifts were
435
estimated that maximize epipolar consistency within an acquisition. The shifts served as
436
basis for an image-based surrogate for cardiac motion and allow to compensate for intra-scan
437
respiratory motion. All figures of merit suggested good performance for both phantom and
438
clinical cases.
439
Future work will investigate extensions of the proposed method, e.g., with respect to motion
440
models, which is of particular interest for experimental imaging trajectories, such as dual axis
441
rotational angiography [37] or saddle trajectories [38]. Finally, the current single-frame sub-
442
traction pipeline relies on sophisticated segmentation and background estimation algorithms
443
that require thorough parameter tuning and exhibit long run times. Although the results
444
reported here are very promising, we believe that a less sophisticated background subtrac-
445
tion pipeline, e.g., based on morphological operations, would drastically ease the transition
446
to clinical application. Possibilities in this regard should, therefore, be investigated.
447
ACKNOWLEDGMENTS
448
The authors gratefully acknowledge funding of DFG MA 4898/3-1 and the Erlangen
449
Graduate School in Advanced Optical Technologies (SAOT) by the German Research Foun-
450
dation (DFG) in the framework of the German excellence initiative, and would like to thank
451
Dr. G¨ unter Lauritsch (Siemens Healthineers) for his valuable comments, and Dr. Christian
31 452
Schlundt and Dr. Michaela Hell (both University Hospital Erlangen) for providing the image
453
data.
CONFLICTS OF INTEREST
454
455
The authors have no relevant conflicts of interest to disclose.
456
[1] H Hetterich, T Redel, G Lauritsch, C Rohkohl, and J Rieber, “New X-ray imaging modal-
457
ities and their integration with intravascular imaging and interventions.” Int. J. Cardiovasc.
458
Imaging 26, 797–808 (2010).
459
[2] N E Green, S Y J Chen, A R Hansgen, J C Messenger, B M Groves, and J D Carroll, “Angio-
460
graphic Views Used for Percutaneous Coronary Interventions: A Three-Dimensional Analysis
461
of Physician-Determined vs. Computer-Generated Views,” Catheter. Cardiovasc. Interv. 64,
462
451–459 (2005).
463
[3] C Blondel, G Malandain, R Vaillant, and N Ayache, “Reconstruction of Coronary Arteries
464
from a Single rotational X-ray projection sequence.” IEEE Trans. Med. Imaging 25, 653–663
465
(2006).
466
467
[4] M Unberath, S Achenbach, R Fahrig, and A Maier, “Exhaustive Graph Cut-based Vasculature Reconstruction,” in Proc. ISBI (IEEE, 2016) pp. 1143–1146.
468
[5] J C Messenger, S Y J Chen, J D Carroll, J E B Burchenal, K Kioussopoulos, and B M
469
Groves, “3D coronary reconstruction from routine single-plane coronary angiograms: Clinical
470
validation and quantitative analysis of the right coronary artery in 100 patients,” Int. J. Card.
471
Imaging 16, 413–427 (2000).
32 472
473
[6] S C ¸ imen, A Gooya, M Grass, and A F Frangi, “Reconstruction of Coronary Arteries from X-ray Angiography: A Review,” Med. Image Anal. (2016).
474
[7] E Hansis, D Sch¨ afer, O D¨ ossel, and M Grass, “Projection-based motion compensation for
475
gated coronary artery reconstruction from rotational x-ray angiograms.” Phys. Med. Biol. 53,
476
3807–20 (2008).
477
[8] C Schwemmer, C Rohkohl, G Lauritsch, K M¨ uller,
and J Hornegger, “Residual motion
478
compensation in ECG-gated interventional cardiac vasculature reconstruction.” Phys. Med.
479
Biol. 58, 3717–37 (2013).
480
481
482
483
484
485
486
487
[9] U Jandt, D Sch¨ afer, M Grass, and V Rasche, “Automatic generation of 3D coronary artery centerlines using rotational X-ray angiography.” Med. Image Anal. 13, 846–58 (2009). [10] B Desjardins and E A Kazerooni, “ECG-gated Cardiac CT,” AJR Am. J. Roentgenol. 182, 993–1010 (2004). [11] J-J Sonke, L Zijp, P Remeijer, and M van Herk, “Respiratory correlated cone beam CT,” Med. Phys. 32, 1176 (2005). [12] T Li, A Koong, and L Xing, “Enhanced 4D cone-beam CT with inter-phase motion model,” Med. Phys. 34, 3688 (2007).
488
[13] M Brehm, P Paysan, M Oelhafen, P Kunz, and M Kachelrieß, “Self-adapting cyclic registra-
489
tion for motion-compensated cone-beam CT in image-guided radiation therapy,” Med. Phys.
490
39, 7603–7618 (2012).
491
[14] M Brehm, P Paysan, M Oelhafen, and M Kachelrieß, “Artifact-resistant motion estimation
492
with a patient-specific artifact model for motion-compensated cone-beam CT,” Med. Phys.
493
40, 1–13 (2013).
33 494
[15] M Brehm, S Sawall, J Maier, S Sauppe,
and M Kachelrieß, “Cardiorespiratory motion-
495
compensated micro-CT image reconstruction using an artifact model-based motion estima-
496
tion,” Med. Phys. 42, 1948–1958 (2015).
497
498
499
[16] A Brost, R Liao, N Strobel, and J Hornegger, “Respiratory motion compensation by modelbased catheter tracking during EP procedures,” Med. Image Anal. 14, 695–706 (2010). [17] G Shechter, F Devernay, E Coste-Mani`ere, A Quyyumi,
and E R McVeigh, “Three-
500
dimensional motion tracking of coronary arteries in biplane cineangiograms.” IEEE Trans.
501
Med. Imaging 22, 493–503 (2003).
502
[18] C Blondel, R Vaillant, G Malandain, and N Ayache, “3D tomographic reconstruction of
503
coronary arteries using a precomputed 4D motion field,” Phys. Med. Biol. 49, 2197–2208
504
(2004).
505
[19] Y Wang, S J Riederer, and R L Ehman, “Respiratory Motion of the Heart: Kinematics and
506
the Implications for the Spatial Resolution in Coronary Imaging,” Magn. Reson. Med. 33,
507
713–719 (1995).
508
[20] G Shechter, C Ozturk, J R Resar, and E R McVeigh, “Respiratory Motion of the Heart From
509
Free Breathing Coronary Angiograms.” IEEE Trans. Med. Imaging 23, 1046–56 (2004).
510
[21] G Shechter, J R Resar, and E R McVeigh, “Displacement and Velocity of the Coronary
511
512
513
514
515
Arteries: Cardiac and Respiratory Motion.” IEEE Trans. Med. Imaging 25, 369–75 (2006). [22] A Aichert, M Berger, J Wang, N Maass, A Doerfler, J Hornegger, and A Maier, “Epipolar Consistency in Transmission Imaging,” IEEE Trans. Med. Imaging 34, 2205–2219 (2015). [23] M Unberath, A Aichert, S Achenbach, and A Maier, “Single-frame Subtraction Imaging,” in Proc. CT Meeting (2016) pp. 89–92.
34 516
517
[24] W P Segars, M Mahesh, T J Beck, E C Frey, and B M W Tsui, “Realistic CT simulation using the 4D XCAT phantom,” Med. Phys. 35, 3800–3808 (2008).
518
[25] A Maier, H G Hofmann, C Schwemmer, J Hornegger, A Keil, and R Fahrig, “Fast simulation
519
of x-ray projections of spline-based surfaces using an append buffer,” Phys. Med. Biol. 57,
520
6193–6210 (2012).
521
522
523
524
525
526
[26] C Tomasi and R Manduchi, “Bilateral filtering for gray and color images,” in Proc. ICCV (IEEE, 1998) pp. 839–846. [27] A F Frangi, W J Niessen, K L Vincken, and M A Viergever, “Multiscale vessel enhancement filtering,” in Proc. MICCAI (Springer, 1998) pp. 130–137. [28] T M Koller, G Gerig, G Szekely, and D Dettwiler, “Multiscale detection of curvilinear structures in 2-D and 3-D image data,” in Proc. ICCV (IEEE, 1995) pp. 864–869.
527
[29] M Berger, C Forman, C Schwemmer, J H Choi, K M¨ uller, A Maier, J Hornegger, and R Fahrig,
528
“Automatic Removal of Externally Attached Fiducial Markers in Cone Beam C-arm CT,” in
529
Proc. BVM (2014) pp. 168–173.
530
531
532
533
[30] A Aach and V Metzler, “Defect interpolation in digital radiography - how object-oriented transform coding helps,” in Proc. SPIE Med. Imaging (2001) pp. 824–835. [31] F J Harris, “On the use of windows for harmonic analysis with the discrete Fourier transform,” Proc. IEEE 66, 51–83 (1978).
534
[32] W R Brody, “Digital subtraction angiography,” IEEE Trans. Nucl. Sci 29, 1176–1180 (1982).
535
[33] M Berger, A Maier, Y Xia, J Hornegger, and R Fahrig, “Motion Compensated Fan-Beam CT
536
by Enforcing Fourier Properties of the Sinogram,” in Proc. CT Meeting (2014) pp. 329–332.
537
[34] A Rodr´ıguez-Molinero, L Narvaiza, J Ruiz, and C G´alvez-Barr´on, “Normal respiratory rate
538
and peripheral blood oxygen saturation in the elderly population,” J. Am. Geriatr. Soc. 61,
35 539
2238–2240 (2013).
540
[35] M Unberath, K Mentl, O Taubmann, S Achenbach, R Fahrig, J Hornegger, and A Maier,
541
“Torsional Heart Motion in Cone-beam Computed Tomography Reconstruction,” in Proc.
542
Fully3D (2015) pp. 651–654.
543
[36] C Debbeler, N Maass, M Elter, F Dennerlein, and T M Buzug, “A new CT rawdata redun-
544
dancy measure applied to automated misalignment correction,” in Proc. Fully3D (2013) pp.
545
264–267.
546
547
548
549
[37] A J P Klein and J A Garcia, “Rotational coronary angiography,” Cardiol. Clin. 27, 395–405 (2009). [38] J D Pack, F Noo, and H Kudo, “Investigation of saddle trajectories for cardiac CT imaging in cone-beam geometry,” Phys. Med. Biol. 49, 2317 (2004).