Pulmonary Arterial Tree Labeling

Labeling the Pulmonary Arterial Tree in CT images for Automatic Quantification of Pulmonary Embolism

Graduation project in collaboration with Dr. Emile A. Hendriks
Information and Communication Theory Group, Technical University Delft.

Ralph Peters, Henk Marquering, Berend Stoel


Pulmonary Embolism (PE) is a condition where one or more of the pulmonary arteries has been (partly) occluded. This occlusion is usually caused by blood clots, thrombi, originating from either the venous circulation or the right side of the heart, but can also be caused for example by tumors that have invaded the circulatory system. The immediate result is partial or complete obstruction of the blood flow to the lungs. The blockade results in a sector of the lungs that is ventilated, but not perfused. When a large embolus occludes a pulmonary artery, the patient suffers acute respiratory distress and may die in a few minutes; a medium-sized embolus may block an artery which supplies blood to a bronchopulmonary segment producing a thrombotic infarct.

CTA Pulmonary EmbolismThe latest standard to diagnose PE is multi-detector computed tomographic angiography (CTA). With this technique, the pulmonary vascular tree is imaged by using computed tomography (CT) in combination with an iodinated contrast agent. This makes the PA tree clearly distinguishable from its surroundings. A PE is then characterized as a dark spot in a bright vessel.


Although CTA has proven to improve the accuracy of PE diagnosis, it also introduces a new problem. Because the radiologist is confronted with a vast amount of anatomical information, analysis of the data is both time consuming and subject to human errors.


Several methods for automated (computer-aided) detection of PE have been proposed. These methods use the fact that in CTA data an arterial occlusion will be characterized by a dark spot inside a bright artery. Recently, a quantification method developed by Qanadli et al. was evaluated by Wu et al. by relating it to patient death in the setting of PE. We decided to follow this approach, since this is the only clinically evaluated method at the moment for quantification of PE. Also, semi-automatic quantification of PE seems to be a relatively unresearched field at the moment.

Qanadli IndexThe quantification index by E.D. Qanadli combines the severity of the stenosis of a pulmonary artery and relative position in the PA tree. To automate this process we therefore need two kinds of information:

  • The size of the occlusion due to the presence of a thrombus.
  • The position of the thrombus in the PA tree.

To come up with this kind of information, first the PA tree needs to be segmented. From this segmented tree volume, the bifurcations can be detected so a tree model can be built. Also, the PE should be detected. Because research has been done in these specific fields (e.g. [5]), we chose to pay less attention to it and roughly implement these steps in the first period of the project.

Quantification of PE in CT data is currently carried out by visual inspection and is associated with significant analysis time and prone to human errors. To realize automatic quantification of PE according to the Qanadli index, automatic extraction of the locations of the 20 arterial segments is required. In this research project, we proposed to come up with a tool, which automatically extracts these segment locations having a segmented binary tree model as input.


Because anatomical knowledge about the segmented branches is required to count the number of affected subsegments, we propose to extract these positions using a matching procedure with a predefined model. To automatically extract the pulmonary tree segments, we propose the following method:

Extract raster points using labeled dataset

  1. Manually label a segmented binary tree according to segment descriptions as described in literature [3], this will result in a point set where every point represents a bifurcation having a certain segment label, the binary tree model also contains parent-child relations between the bifurcation points
  2. Split into left and right tree, stretch the bifurcation points so they fit in a 1x1x1 box using linear scaling, define a certain sampling raster (e.g. 20x20x20), for each raster point: find the nearest bifurcation point and use this label

Figure 0

Use raster points of several datasets to define segment reference points

  1. Use multiple rasters from a number of manually labeled datasets
  2. Find intersection for each segment
  3. Find relative reference points by averaging the raster points in each intersected segment

Figure 1

Use segment reference points to extract segments from an unlabeled dataset

  1. The extraction has an unlabeled dataset as input, this is a segmented binary tree containing the bifurcation positions of the pulmonary arterial tree with parent-child relations
  2. First split into left and right tree using parent-child relations
  3. Place found reference points into unlabeled points set (again using linear scaling), for each reference point find the nearest unlabeled point and label this point
  4. by labeling a parent points to the tree root using the parent-child relations, the origin of each segment can be found, using these segment origins the entire tree can be labeled



Below the results of the leave-one-out tests is presented, using the previously described method on ten datasets. We added the results of an extra feature in the method: extracting the left and right tree separately, and extracting the lower segments of the left and right tree separately (because these segments generally have a common trunk). Extracting the lower segments separately will add robustness to anatomical variations, especially in the Z-direction (perpendicular to the slice).

Figure 3


Ralph Peters, Emile A. Hendriks, Henk Marquering, B.C. Stoel, H. Dogan, A. de Roos, J.H.C. Reiber, "Labeling the pulmonary arterial tree in CT images for automatic quantification of pulmonary embolism", Proceedings of SPIE - Volume 6514 - Medical Imaging -Computer-Aided Diagnosis, San Diego, no. 6514, February 2007.


For further information, please contact:
B.C. Stoel, PhD.
Division of Image Processing
Department of Radiology, 1-C2S
Leiden University Medical Center
P.O. Box 9600
2300 RC Leiden
The Netherlands
Tel. +31 (0)71 526 1911
Fax. +31 (0)71 526 6801
e-mail: B.C.Stoel@lumc.nl