System and method for computer aided detection of pulmonary embolism in tobogganing in CT angiography

1. A method for detecting pulmonary embolisms in computed tomographic (CT) angiography using a processor, said method comprising the steps of:

providing a digitized CT image acquired from the inside of a pulmonary vessel, said image comprising a plurality of intensities corresponding to a 3-dimensional grid comprising a first plurality of voxels;

for each of the first plurality of voxels in said image,

extracting a first pulmonary embolism (PE) candidate and first PE boundary; and

for each of a second plurality of voxels within said first PE boundary,

selecting a voxel of the second plurality of voxels;

extracting a subsequent PE candidate and subsequent PE boundary from said selected voxel;

merging said subsequent PE candidate with said first PE candidate; and

merging said subsequent PE boundary with said first PE boundary;

wherein extracting a first or second pulmonary embolism (PE) candidate and first or second PE boundary from each of a first or second plurality of voxels, respectively comprises:

tobogganing said voxel to a first concentration voxel:

initializing a cluster with said first concentration voxel, and assigning to an active list all neighboring voxels whose intensity is indicative of being part of an embolism;

removing a voxel with a minimal intensity from said active list;

tobogganing said removed voxel to a second concentration voxel, wherein if said second concentration voxel is part of said cluster, assigning neighbors of said removed voxel to said active list, otherwise assigning said removed voxel to a boundary of said cluster; and

repeating said steps of removing a voxel from said active list and tobogganing said removed voxel to a second concentration voxel until said active list is empty, wherein said cluster is the pulmonary embolism candidate; and

wherein each of the foregoing steps is performed using a processor.

2. The method of claim 1 , wherein tobogganing a voxel comprises finding a concentration voxel from among said voxel and its neighboring voxels that has a minimum intensity.

3. The method of claim 1 , further comprising adding said removed voxel to said cluster, when said second concentration voxel is part of said cluster.

4. The method of claim 1 , wherein an intensity indicative of being part of an embolism is in the range of between about −50 Hounsfield Units (HU) and about 100 HU.

5. The method of claim 1 , further comprising classifying said pulmonary embolism candidate as a true candidate or a false candidate by calculating an image-based feature vector x for said candidate and comparing said feature vector to a classification function w T x+b which divides a feature space into a false candidate space and a true candidate space.

6. The method of claim 5 , wherein said classifier function coefficients w and intercepts b are found by solving

min

w

,

ξ

⁢

γ

⁢



w



1

+

∑

j

=

1

m

⁢

⁢

(

∑

i

∈

I

j

⁢

⁢

λ

i

⁢

ξ

i

)

+

∑

i

∈

I

-

⁢

⁢

ξ

i

,

⁢

s

.

t

.

⁢

w

T

⁢

x

i

+

b

≥

1

-

ξ

i

,

i

∈

I

+

,

I

+

=

⋃

m

j

=

1

⁢

I

j

w

T

⁢

x

i

+

b

≤

-

1

+

ξ

i

,

i

∈

I

-

,

⁢

ξ

i

≥

0

,

i

=

1

,

…

⁢

,

l

,

⁢

∑

i

∈

I

j

⁢

⁢

λ

i

=

1

,

λ

i

≥

0

,

i

∈

I

j

,

j

=

1

,

…

⁢

,

m

,

for a training set of true and false pulmonary embolism candidates, wherein I + and I − are index sets containing indices of true candidates that intersect with PEs and false candidates that do not intersect with PEs, respectively, and I j is an index set of the candidates that intersect with a j th PE, j=1, . . . , m.

7. A program storage device readable by a computer, tangibly embodying a program of instructions executable by the computer to perform the method steps for detecting pulmonary embolisms in computed tomographic (CT) angiography, said method comprising the steps of:

providing a digitized CT image acquired from the inside of a pulmonary vessel, said image comprising a plurality of intensities corresponding to a 3-dimensional grid comprising a first plurality of voxels;

for each of the first plurality of voxels in said image,

extracting a first pulmonary embolism (PE) candidate and first PE boundary; and

for each of a second plurality of voxels within said first PE boundary,

selecting a voxel of the second plurality of voxels;

extracting a subsequent PE candidate and subsequent PE boundary from said selected voxel;

merging said subsequent PE candidate with said first PE candidate; and

merging said subsequent PE boundary with said first PE boundary

wherein extracting a first or second pulmonary embolism (PE) candidate and first or second PE boundary from each of a first or second plurality of voxels, respectively comprises:

tobogganing said voxel to a first concentration voxel;

initializing a cluster with said first concentration voxel, and assigning to an active list all neighboring voxels whose intensity is indicative of being part of an embolism;

removing a voxel with a minimal intensity from said active list;

tobogganing said removed voxel to a second concentration voxel, wherein if said second concentration voxel is part of said cluster, assigning neighbors of said removed voxel to said active list, otherwise assigning said removed voxel to a boundary of said cluster; and

repeating said steps of removing a voxel from said active list and tobogganing said removed voxel to a second concentration voxel until said active list is empty, wherein said cluster is the pulmonary embolism candidate.

8. The computer readable program storage device of claim 7 , wherein tobogganing a voxel comprises finding a concentration voxel from among said voxel and its neighboring voxels that has a minimum intensity.

9. The computer readable program storage device of claim 7 , the method further comprising adding said removed voxel to said cluster, when said second concentration voxel is part of said cluster.

10. The computer readable program storage device of claim 7 , wherein an intensity indicative of being part of an embolism is in the range of between about −50 Hounsfield Units (HU) and about 100 HU.

11. The computer readable program storage device of claim 7 , the method further comprising classifying said pulmonary embolism candidate as a true candidate or a false candidate by calculating an image-based feature vector x for said candidate and comparing said feature vector to a classification function w T x+b which divides a feature space into a false candidate space and a true candidate space.

12. The computer readable program storage device of claim 11 , wherein said classifier function coefficients w and intercepts b are found by solving

min

w

,

ξ

⁢

γ

⁢



w



1

+

∑

j

=

1

m

⁢

⁢

(

∑

i

∈

I

j

⁢

⁢

λ

i

⁢

ξ

i

)

+

∑

i

∈

I

-

⁢

⁢

ξ

i

,

⁢

s

.

t

.

⁢

w

T

⁢

x

i

+

b

≥

1

-

ξ

i

,

i

∈

I

+

,

I

+

=

⋃

m

j

=

1

⁢

I

j

w

T

⁢

x

i

+

b

≤

-

1

+

ξ

i

,

i

∈

I

-

,

⁢

ξ

i

≥

0

,

i

=

1

,

…

⁢

,

l

,

⁢

∑

i

∈

I

j

⁢

⁢

λ

i

=

1

,

λ

i

≥

0

,

i

∈

I

j

,

j

=

1

,

…

⁢

,

m

,

for a training set of true and false pulmonary embolism candidates, wherein I + and I − are index sets containing indices of true candidates that intersect with PEs and false candidates that do not intersect with PEs, respectively, and I j is an index set of the candidates that intersect with a j th PE, j=1, . . . , m.