Limits...
Segmentation of Pulmonary Vascular Trees from Thoracic 3D CT Images.

Shikata H, McLennan G, Hoffman EA, Sonka M - Int J Biomed Imaging (2009)

Bottom Line: A quantitative validation was performed with more than 1000 manually identified points selected from inside the vessel segments to assess true positives (TPs) and 1000 points randomly placed outside of the vessels to evaluate false positives (FPs) in each case.On average, for both the high and low volume lung images, 99% of the points was properly marked as vessel and 1% of the points were assessed as FPs.Our hybrid segmentation algorithm provides a highly reliable method of segmenting the combined pulmonary venous and arterial trees which in turn will serve as a critical starting point for further quantitative analysis tasks and aid in our overall goal of establishing a normative atlas of the human lung.

View Article: PubMed Central - PubMed

Affiliation: Iowa Institute for Biomedical Imaging, The University of Iowa, Iowa City, IA 52242, USA.

ABSTRACT
This paper describes an algorithm for extracting pulmonary vascular trees (arteries plus veins) from three-dimensional (3D) thoracic computed tomographic (CT) images. The algorithm integrates tube enhancement filter and traversal approaches which are based on eigenvalues and eigenvectors of a Hessian matrix to extract thin peripheral segments as well as thick vessels close to the lung hilum. The resultant algorithm was applied to a simulation data set and 44 scans from 22 human subjects imaged via multidetector-row CT (MDCT) during breath holds at 85% and 20% of their vital capacity. A quantitative validation was performed with more than 1000 manually identified points selected from inside the vessel segments to assess true positives (TPs) and 1000 points randomly placed outside of the vessels to evaluate false positives (FPs) in each case. On average, for both the high and low volume lung images, 99% of the points was properly marked as vessel and 1% of the points were assessed as FPs. Our hybrid segmentation algorithm provides a highly reliable method of segmenting the combined pulmonary venous and arterial trees which in turn will serve as a critical starting point for further quantitative analysis tasks and aid in our overall goal of establishing a normative atlas of the human lung.

No MeSH data available.


Related in: MedlinePlus

Segmentation results by thresholding the filter output with a fixed value (−600 [H.U.]). (a) Original CT image visualized using a window of 1500 H.U. at the level of  −400 H.U. (b) A fixed  is used for all voxels. (c) (2) is used to determine S. In (b), nonvessel regions around a junction are also extracted as vessels whereas segmented regions reside within the visible vessels in (c). It should be noted that while vessel segments do not appear to be connected in this image, they are connected in the neighboring slices. In the event that they were not connected, the local disconnections would be fixed by the tracking process described in Section 2.2.
© Copyright Policy - open-access
Related In: Results  -  Collection


getmorefigures.php?uid=PMC2801012&req=5

fig3: Segmentation results by thresholding the filter output with a fixed value (−600 [H.U.]). (a) Original CT image visualized using a window of 1500 H.U. at the level of −400 H.U. (b) A fixed is used for all voxels. (c) (2) is used to determine S. In (b), nonvessel regions around a junction are also extracted as vessels whereas segmented regions reside within the visible vessels in (c). It should be noted that while vessel segments do not appear to be connected in this image, they are connected in the neighboring slices. In the event that they were not connected, the local disconnections would be fixed by the tracking process described in Section 2.2.

Mentions: (1)F(x)=max σf∈S−σf2λ2I(x), where σf is the standard deviation of a Gaussian function convoluted with an image so as to take second derivatives in the volume coordinate, and I(x) is the intensity value at the point. S is a discrete set of σf for multiscale integration. This equation only takes into account criterion 1. Thus, it may also enhance blob structures, typically image noise as well as cylindrical structures. According to our experience, in thoracic CT images, the noise will not be enhanced to the same degree as are vessels since contrast between parenchymal background and pulmonary vessels is relatively high compared to the noise. Most of the visible thin vessel segments exhibit intensity over −700 H.U. (Hounsfield Units) while lung parenchyma is typically between −800 and −900 H.U. This results in a 100 to 200 H.U. difference. Since Gaussian noise with standard deviation σ = 20 can be considered typical for CT images [8], a difference greater than 100 H.U. is not caused by noise. The noise can be eliminated by postprocessing discussed in Section 2.3. F(x) takes a maximum value when σf is the closest to the radius of a target segment among other σf in the range S. Therefore, S should include appropriate values that cover all the radii of the vessel segments in the lung. However, when S contains a wide range of values, and the filter output is calculated at a nonvessel point close to a junction or multiple thin segments are close to each other, F(x) gets larger than it should. This is caused by a Gaussian function with a large σf that excessively smoothes the region. Figure 3(b) shows the result when a fixed range of S was used for all voxels in the image. The filter overly extracted nonvessel regions around the junction. In order to avoid the inappropriate enhancement, large σf should be included in S only when it needs to be used for the detection of a thick segment. This requires a priori knowledge of the thick vessels. In thoracic images, thick vessels can easily be extracted by a simple intensity-based thresholding. We applied thresholding with a fixed value to obtain thick segments, and a discrete distance transform is applied to estimate an approximate radius r. By using a distance transform value d, the radius r is estimated as [voxel]. We used −600 [H.U.] for both TLC and FRC scans as the threshold value. Using the radius information, range S is determined depending on the distance d as follows:


Segmentation of Pulmonary Vascular Trees from Thoracic 3D CT Images.

Shikata H, McLennan G, Hoffman EA, Sonka M - Int J Biomed Imaging (2009)

Segmentation results by thresholding the filter output with a fixed value (−600 [H.U.]). (a) Original CT image visualized using a window of 1500 H.U. at the level of  −400 H.U. (b) A fixed  is used for all voxels. (c) (2) is used to determine S. In (b), nonvessel regions around a junction are also extracted as vessels whereas segmented regions reside within the visible vessels in (c). It should be noted that while vessel segments do not appear to be connected in this image, they are connected in the neighboring slices. In the event that they were not connected, the local disconnections would be fixed by the tracking process described in Section 2.2.
© Copyright Policy - open-access
Related In: Results  -  Collection

Show All Figures
getmorefigures.php?uid=PMC2801012&req=5

fig3: Segmentation results by thresholding the filter output with a fixed value (−600 [H.U.]). (a) Original CT image visualized using a window of 1500 H.U. at the level of −400 H.U. (b) A fixed is used for all voxels. (c) (2) is used to determine S. In (b), nonvessel regions around a junction are also extracted as vessels whereas segmented regions reside within the visible vessels in (c). It should be noted that while vessel segments do not appear to be connected in this image, they are connected in the neighboring slices. In the event that they were not connected, the local disconnections would be fixed by the tracking process described in Section 2.2.
Mentions: (1)F(x)=max σf∈S−σf2λ2I(x), where σf is the standard deviation of a Gaussian function convoluted with an image so as to take second derivatives in the volume coordinate, and I(x) is the intensity value at the point. S is a discrete set of σf for multiscale integration. This equation only takes into account criterion 1. Thus, it may also enhance blob structures, typically image noise as well as cylindrical structures. According to our experience, in thoracic CT images, the noise will not be enhanced to the same degree as are vessels since contrast between parenchymal background and pulmonary vessels is relatively high compared to the noise. Most of the visible thin vessel segments exhibit intensity over −700 H.U. (Hounsfield Units) while lung parenchyma is typically between −800 and −900 H.U. This results in a 100 to 200 H.U. difference. Since Gaussian noise with standard deviation σ = 20 can be considered typical for CT images [8], a difference greater than 100 H.U. is not caused by noise. The noise can be eliminated by postprocessing discussed in Section 2.3. F(x) takes a maximum value when σf is the closest to the radius of a target segment among other σf in the range S. Therefore, S should include appropriate values that cover all the radii of the vessel segments in the lung. However, when S contains a wide range of values, and the filter output is calculated at a nonvessel point close to a junction or multiple thin segments are close to each other, F(x) gets larger than it should. This is caused by a Gaussian function with a large σf that excessively smoothes the region. Figure 3(b) shows the result when a fixed range of S was used for all voxels in the image. The filter overly extracted nonvessel regions around the junction. In order to avoid the inappropriate enhancement, large σf should be included in S only when it needs to be used for the detection of a thick segment. This requires a priori knowledge of the thick vessels. In thoracic images, thick vessels can easily be extracted by a simple intensity-based thresholding. We applied thresholding with a fixed value to obtain thick segments, and a discrete distance transform is applied to estimate an approximate radius r. By using a distance transform value d, the radius r is estimated as [voxel]. We used −600 [H.U.] for both TLC and FRC scans as the threshold value. Using the radius information, range S is determined depending on the distance d as follows:

Bottom Line: A quantitative validation was performed with more than 1000 manually identified points selected from inside the vessel segments to assess true positives (TPs) and 1000 points randomly placed outside of the vessels to evaluate false positives (FPs) in each case.On average, for both the high and low volume lung images, 99% of the points was properly marked as vessel and 1% of the points were assessed as FPs.Our hybrid segmentation algorithm provides a highly reliable method of segmenting the combined pulmonary venous and arterial trees which in turn will serve as a critical starting point for further quantitative analysis tasks and aid in our overall goal of establishing a normative atlas of the human lung.

View Article: PubMed Central - PubMed

Affiliation: Iowa Institute for Biomedical Imaging, The University of Iowa, Iowa City, IA 52242, USA.

ABSTRACT
This paper describes an algorithm for extracting pulmonary vascular trees (arteries plus veins) from three-dimensional (3D) thoracic computed tomographic (CT) images. The algorithm integrates tube enhancement filter and traversal approaches which are based on eigenvalues and eigenvectors of a Hessian matrix to extract thin peripheral segments as well as thick vessels close to the lung hilum. The resultant algorithm was applied to a simulation data set and 44 scans from 22 human subjects imaged via multidetector-row CT (MDCT) during breath holds at 85% and 20% of their vital capacity. A quantitative validation was performed with more than 1000 manually identified points selected from inside the vessel segments to assess true positives (TPs) and 1000 points randomly placed outside of the vessels to evaluate false positives (FPs) in each case. On average, for both the high and low volume lung images, 99% of the points was properly marked as vessel and 1% of the points were assessed as FPs. Our hybrid segmentation algorithm provides a highly reliable method of segmenting the combined pulmonary venous and arterial trees which in turn will serve as a critical starting point for further quantitative analysis tasks and aid in our overall goal of establishing a normative atlas of the human lung.

No MeSH data available.


Related in: MedlinePlus