System and method for quasi-real-time ventricular measurements from M-mode echocardiogram
View Patent ↗A method for measuring ventricular dimensions from M-mode echocardiograms, includes providing a digitized M-mode echocardiogram image, running a plurality of local classifiers, where each local classifier trained to detect a landmark on either an end-diastole (ED) line or an end-systole (ES) line in the image, recording all possible landmarks detected by the classifiers, where a search range in an N-dimensional parameter space defined by the landmarks for each dimension is reduced to a union of subsets, where each dimension of the parameter space corresponds a landmark, for each combination of possible landmarks, checking if an order of the landmarks is consistent with a known ordering of the landmarks, and if the order is consistent, running a global detector on each consistent combination of landmarks to find a landmark combination with a highest detection probability as a confirmed landmark detection, where the landmarks are used for measuring ventricular dimensions.
1. A computer implemented method for measuring ventricular dimensions from M-mode echocardiograms, the method performed by the computer comprising the steps of:
providing to a computer a digitized M-mode echocardiogram image comprising a plurality of intensities associated with a 2-dimensional grid;
running a plurality of local classifiers, wherein each local classifier trained to detect a landmark on either an end-diastole (ED) line or an end-systole (ES) line in said image, and recording all possible landmarks detected by said classifiers, wherein a search range in an N-dimensional parameter space defined by said landmarks for each dimension is reduced into a union of subsets containing said possible landmark detection, wherein each dimension of said parameter space corresponds to one of the plurality of landmarks;
for each combination of possible landmarks, checking if an order of the possible landmarks is consistent with a known ordering of said landmarks; and
if said order is consistent, running a global detector on each consistent combination of possible landmarks to find a possible landmark combination with a highest detection probability as a confirmed landmark detection, wherein said landmarks are used for measuring ventricular dimensions.
2. The method of claim 1 , wherein there are 5 landmarks on the ED line and 4 landmarks on the ES line.
3. The method of claim 1 , further comprising:
smoothing a landmark probability response line output by a local classifier to find all local maxima;
ranking the local maxima based on their response magnitudes; and
for each local maxima, if a minimum distance of a local maximum with respect to another local maximum in a mode point set is greater than a pre-determined threshold, then said local maximum is added to the mode point set, wherein said mode point set forms a combination of possible landmarks, and wherein said mode point set is initialized to local maximum having a highest response from said local classifiers.
4. The method of claim 1 , further comprising:
warping locations {y 1 , . . . , y N } of said possible landmarks along a line to coincide with canonical locations {z 1 , . . . , z N } of known landmarks by solving for a function
f
(
z
)
=
∑
n
=
1
,
…
,
N
c
n
ϕ
(
z
-
z
n
)
+
z
·
d
,
where φ(r)=r 2 log(r) and {c 1 , c 2 , . . . , c N , d} are coefficients, wherein y n =f(z n ).
5. The method of claim 4 , wherein the coefficients {c 1 , c 2 , . . . , c N , d} are determined by solving
[
y
1
y
2
⋮
y
N
d
]
=
[
0
ϕ
(
z
1
-
z
2
)
…
ϕ
(
z
1
-
z
N
)
z
1
ϕ
(
z
2
-
z
1
)
0
…
ϕ
(
z
2
-
z
N
)
z
2
⋮
⋮
⋮
⋮
⋮
ϕ
(
z
N
-
z
1
)
ϕ
(
z
N
-
z
2
)
…
0
z
N
0
0
…
0
1
]
[
c
1
c
2
⋮
c
N
d
]
.
6. The method of claim 1 , wherein each of said local classifiers and said global classifier are binary decision trees trained as probabilistic boosting trees wherein each tree node is a strong classifier that combines multiple weak classifiers trained on images with pre-identified landmarks.
7. The method of claim 6 , wherein said weak classifiers are modeled by Haar wavelet features that represent image features.
8. A computer implemented method for measuring ventricular dimensions from M-mode echocardiograms, the method performed by the computer comprising the steps of:
providing to a computer a digitized M-mode echocardiogram image comprising a plurality of intensities associated with a 2-dimensional grid;
searching for a set of N landmarks on either an end-diastole (ED) line or an end-systole (ES) line in said image that maximizes a detection probability
p
(
O
|
I
(
y
1
,
y
2
,
…
,
y
N
)
)
=
∏
n
=
1
,
…
N
p
n
L
(
O
|
I
(
y
n
)
)
·
p
G
(
O
|
I
(
y
1
,
y
2
,
…
,
y
N
)
)
,
wherein (y 1 , y 2 , . . . , y N ) are landmark locations in said image, I(y 1 , y 2 , . . . , y N ) is an image patch extracted from said image using the parameter (y 1 , y 2 , . . . , y N ), p(O|I(y 1 , y 2 , . . . , y N )) is a probability of being a target landmark, I(y n ) is a local image patch defined by landmark y n , p n L (O|I(y n ))) is a local classifier for landmark y n , and p G (O|I(y 1 , y 2 , . . . , y N )) is a global detection probability of detecting said N landmarks simultaneously, wherein said landmarks are used for measuring ventricular dimensions.
9. The method of claim 8 , wherein running the local detectors reduces an N-dimensional parameter space defined by said landmarks into a union of subsets containing possible landmark detections detected by said local classifiers by filtering those that fail to be detected by the local detector, wherein each dimension of said parameter space corresponds to one of the set of landmarks, and further comprising:
for each combination of possible landmarks, checking if an order of the possible landmarks is consistent with a known ordering of said landmarks; and
running said global detector on each consistent combination of possible landmarks to find a possible landmark combination with a highest detection probability as a confirmed landmark detection.
10. A non-transitory program storage device readable by a computer, tangibly embodying a program of instructions executable by the computer to perform the method steps for measuring ventricular dimensions from M-mode echocardiograms, said method comprising the steps of:
providing a digitized M-mode echocardiogram image comprising a plurality of intensities associated with a 2-dimensional grid;
running a plurality of local classifiers, wherein each local classifier trained to detect a landmark on either an end-diastole (ED) line or an end-systole (ES) line in said image, and recording all possible landmarks detected by said classifiers, wherein a search range in an N-dimensional parameter space defined by said landmarks for each dimension is reduced into a union of subsets containing said possible landmark detection, wherein each dimension of said parameter space corresponds to one of the plurality of landmarks;
for each combination of possible landmarks, checking if an order of the possible landmarks is consistent with a known ordering of said landmarks; and
if said order is consistent, running a global detector on each consistent combination of possible landmarks to find a possible landmark combination with a highest detection probability as a confirmed landmark detection, wherein said landmarks are used for measuring ventricular dimensions.
11. The computer readable program storage device of claim 10 , wherein there are 5 landmarks on the ED line and 4 landmarks on the ES line.
12. The computer readable program storage device of claim 10 , the method further comprising:
smoothing a landmark probability response line output by a local classifier to find all local maxima;
ranking the local maxima based on their response magnitudes; and
for each local maxima, if a minimum distance of a local maximum with respect to another local maximum in a mode point set is greater than a pre-determined threshold, then said local maximum is added to the mode point set, wherein said mode point set forms a combination of possible landmarks, and wherein said mode point set is initialized to local maximum having a highest response from said local classifiers.
13. The computer readable program storage device of claim 10 , the method further comprising:
warping locations {y 1 , . . . , y N } of said possible landmarks along a line to coincide with canonical locations {z 1 , . . . , z N } of known landmarks by solving for a function
f
(
z
)
=
∑
n
=
1
,
…
,
N
c
n
ϕ
(
z
-
z
n
)
+
z
·
d
,
where φ(r)=r 2 log(r) and {c 1 , c 2 , . . . , c N , d} are coefficients, wherein y n =f(z n ).
14. The computer readable program storage device of claim 13 , wherein the coefficients {c 1 , c 2 , . . . , c N , d} are determined by solving
[
y
1
y
2
⋮
y
N
d
]
=
[
0
ϕ
(
z
1
-
z
2
)
…
ϕ
(
z
1
-
z
N
)
z
1
ϕ
(
z
2
-
z
1
)
0
…
ϕ
(
z
2
-
z
N
)
z
2
⋮
⋮
⋮
⋮
⋮
ϕ
(
z
N
-
z
1
)
ϕ
(
z
N
-
z
2
)
…
0
z
N
0
0
…
0
1
]
[
c
1
c
2
⋮
c
N
d
]
.
15. The computer readable program storage device of claim 10 , wherein each of said local classifiers and said global classifier are binary decision trees trained as probabilistic boosting trees wherein each tree node is a strong classifier that combines multiple weak classifiers trained on images with pre-identified landmarks.
16. The method of claim 15 , wherein said weak classifiers are modeled by Haar wavelet features that represent image features.