Introduction
Introduction
Clinical diagnosis of pulmonary embolism (PE) is often difficult or even impossible and remains a diagnostic challenge in modern medicine [1 1 2 3 4 5 6
Subsecond spiral CT (0.75 s) is associated with improved clarity and diminished motion artifacts on lung and especially mediastinal structures when compared with 1-second spiral CT [7 8
As known from pulmonary angiography, perfusion defects allow the direct assessment of the extent of PE. Therefore, post-contrast CT data sets of the lung are of high interest as well. Dynamic scanning of the entire lung is not yet possible, since detector aperture is still too small. Nevertheless, parenchymal enhancement is directly correlated to functional parameters such as lung perfusion. Thus, the goal of our study was to extract additional information related to parenchymal enhancement using a new dedicated image processing technique.
Materials and Methods
Materials and Methods
The image processing was performed on clinical data sets, acquired from MSCT chest examinations for clinical suspicion of PE. Six patients (two male, four female) with a mean age of 65 years (48-72 y, ± 8.2 y) were examined. All were clinically stable, regarding their cardiac and respiratory status. Non-ionic contrast media (Ultravist 370, Schering, Berlin, Germany) was applied intravenously using a double power injector (CT 9000 Digital Injection System; Liebel-Flarsheim, Cincinnatti, OH). Flow parameters were 3 cc/s, with a total amount of 120 cc, followed by a saline chaser bolus (3 cc/s, 30 cc). Start delay of all examinations was 30 s in cranio-caudad direction. Imaging was performed on a commercially available MSCT (Somatom Volume Zoom; Siemens, Forchheim, Germany). Scan parameters were 140 kV and 100 mAs, using a rotation time of 0.5 sec, thin collimation of 4 × 1 mm and a table speed of 7 mm (pitch: 1.75). Therefore, the entire chest was examined within a 19.7-22.1 s (m = 21.0 ± 0.31) breathhold. No additional late phase scanning was required.
First, axial images were reconstructed twice. Two sets were reconstructed with a slice thickness of 5 mm, using a standard soft tissue kernel and an edge-enhanced kernel for lung window settings. Technical quality of the data sets was evaluated first. Therefore, Hounsfield Units (HU) within the main pulmonary vessels were characterized with ROI-methodology. Radiological diagnoses regarding PE were established on these axial slices on hard copies as well as on an additionally generated set of thin collimated slices (slice thicknesseff. 1.25 mm, reconstruction increment 0.8 mm, standard soft tissue kernel) in cine mode view on a workstation (Wizard, Siemens). For the patients studied, 292-354 axial source images were calculated (325 ± 19.3). Complete or incomplete occlusion of pulmonary arteries as direct visualization of emboli due to filling defects were used as signs of PE.
The image processing algorithm for visualization of post-contrast parenchymal enhancement was structured into 5 steps: segmentation, vessel cutting, adaptive filtering, color-coding and overlay with the original images.
Contour definition by identifying the lung using threshold based contour finding
In a first step, a binary mask of the image was derived by identifying lung areas and non-lung areas. A threshold based contour finder was used for segmentation, as lung tissue (low HU) is generally enclosed by high HU areas (pleura, chest wall). The threshold HU value separated both HU ranges and defined the contour of the lung. A typical value for the segmentation threshold was -300 HU.
To exclude pleural walls, 5 layers of pixels were removed at the border of segmented lung areas by applying the morphological operation of erosion five times to the binary segmentation mask, using the four connected neighbours as structuring element.
All following steps were performed on the extracted lung-areas, only.
Vessel cutting
Major vascular structures and airways were removed by HU range selection to prepare images for subsequent filtering.
A lower threshold HUB and an upper threshold HUV were specified, pixels below HUB were identified as airways, pixels above HUV as vessels. For optimal image visualization, vessel cutting had to be balanced with preserving as many lung pixels as possible. The algorithm removed all pixels below HUB and adjusted HUV to reach a maximum share of 28 % of removed pixels.
Adaptive filtering
The segmented dataset was reformarted by linear interpolation to obtain a dataset with isotropic voxel spacing. An adaptive sliding mean value filter was applied, using a spherical 3D kernel with a diameter of 5 mm. In most datasets, this corresponded to 7 pixels. In this process, up to 7 adjacent slices were combined. Lost pixels due to vessel cutting were replaced to a specified extent, by the average of pixel value from their 3D environment. Removed pixels did not contribute to the averaging density values.
Color coding
To facilitate visualization of parenchymal enhancement, the resulting image was mapped onto a spectral color scale. Mapping was controlled by center and width, analogous to gray scale image mapping. As subtraction technique from native data was not possible, no baseline was defined. A heuristic approach analyzed the parenchymal histogram and determined the window parameters automatically. The algorithm used the mean value of the processed parenchymal pixels as the center value, the width range was fixed to 100 HU.
Image fusion with the original image
The resulting color-encoded parenchymal images were overlayed onto the original CT images, as these CT source images were crucial for spatial orientation in the dataset. For overlay, all non parenchymal pixels were replaced by the original pixels of the respective slice position and displayed in the usual CT gray scale presentation. In addition, manual interactive windowing of the gray and the colored parts of the image by the user was possible.
A pictorial overview of the image processing algorithm is given in Fig. [1
Fig. 1 Image processing algorithm: (a ) Original image, (b ) Contour detection, (c ) Split-image, lung/non-lung, (d ) Removal of pleura walls, (e ) Removal of vessels and bronchi, (f ) 3D adaptive filtering on image-stack, (g ) Color-coding and overlay to original image
Results
Results
Density distribution (Hounsfield Units [HU]) in the pulmonary arteries was measured using ROI methodology. Mean HU values in the pulmonary trunk were 326 ± 50.0 HU, 317 ± 56.6 HU in the left main pulmonary artery and 307 ± 55.3 HU in the right main pulmonary artery. There fore, a homogeneous distribution of contrast material was achieved in all six patients. Also the HU values of the patients with proven PE were comparable to the healthy patients in this respect.
In all six patients, mean density values on CT perfusion weighted color maps were comparable, ranging from -860 HU to -915 HU (m = -892 ± 17.3 HU).
In the four patients with normal CT scans in axial scanning, color-coded display of lung parenchyma showed quite a homogeneous appearance of density values, displayed in bright green and partly yellow colors (Fig. [2 a-c ). Anatomical details, such as lung fissures, were sharply delineated and allowed anatomical orientation, even on sagittal and coronal images.
In the patients with proven central and peripheral PE, filling defects on CTA corresponded to areas of decreased densities, as arterial inflow into the lung parenchyma was impaired significantly in these lung segments. These areas were predominantly displayed in violet and dark blue colors (Fig. [3 a-c ).
Fig. 2 (a ) Axial image at the level of the left atrium after intravenous administration of contrast medium. Complete enhancement of all lung vessels. No perfusion deficits are detected on lung densitometry. Note the gravity dependent gradient shift in ventro-dorsal direction. (b ) Sagittal view with homogeneous color coding of the entire lung beside the gravity dependent gradient. No local areas with increased or decreased lung density values. The major septum can be delineated (arrows). (c ) In coronal plane, homogeneous density distribution is depicted.
Fig. 3 (a ) Occluding thrombi in the intermediate, middle and lower lobe artery on the right side (arrows). The corresponding lung segments show decreased lung density values and are sharply delineated to the adjacent lung parenchyma. Subsegmental areas of decreased density are seen in the right middle lobe (lateral segment). Note also areas with increased density values (red color coded). Here, a ground glass pattern was detected in lung window settings. (b ) In sagittal view, peripheral deficits are delineated in the right lower lobe as well as in the right upper lobe (apical segment). The latter shows triangular shaped perfusion deficits, characteristic of peripheral pulmonary embolism. (c ) In the coronal view at the level of the left atrium additional (sub)segmental areas with decreased densities are seen in the left upper (apicoposterior segment) and the left lower lobe corresponding to the occluded vessels on CTA of the pulmonary arteries. Note the thrombus in the apicoposterior left lung artery (black arrow).
Discussion
Discussion
Visualization of lung perfusion as an adjunct to CTA seems to be an attractive technique to improve the diagnostic accuracy of PE diagnostics. Especially, as segmental and subsegmental emboli are detected more often, when thin collimation spiral CT examinations are performed [3 6 9
Thin-collimation, sub-second spiral CT leads to improved spatial resolution and consequently increased analyzable subsegmental pulmonary arteries [6 3 2 3
Wintersperger et al. [10 11
Depiction of four slices within one gantry rotation allows a maximum volume coverage of 20 mm for dynamic MSCT studies at the moment. A subtraction technique, as known from angiography, is technically not yet feasible using MSCT.
Groell et al [12
With MSCT, scanning of the entire chest within a single breathhold is technically feasible, even with thin collimation protocols. These protocols are necessary for generating nearly isotropic voxels with adequate delineation in z-direction. This is mandatory for displaying the acquired CT data not only in axial, but also in coronal and sagittal view directions [13
Coulden et al. [14
In the postprocessing method presented, data sets are displayed by color-coding in axial, coronal as well as in sagittal view directions. As the parenchymal histogram and the window parameters are determined automatically, no further (subjective) windowing is needed. Information of the lung perfusion can be obtained by careful analysis of anatomic detail in all three directions. Occluded vessels led to a decrease of blood flow in the corresponding lung parenchyma, which was displayed in dark blue and violet colors. In combination with coronal and sagittal orientations, the lung segments affected were easily delineated.
This methodology will have limitations in patients with underlying emphysema due to their inappropriate re-distribution. Also interpretation of areas with increased density values remains challenging, as several conditions may lead to this finding. In the patients with proven PE, dystelectatic areas led to increasing lung densities, which therefore have to be carefully analyzed on the source images. Other additional findings, such as pleural effusion, ground glass and mosaic attenuation patterns as well as atelectases for several reasons might also have effects on the methodology presented.
Perfusion weighted color maps utilizing MSCT may serve as an additional imaging technique derived from data sets of the entire chest. However, a comparison with ventilation/perfusion scintigraphy and a larger number of patients are necessary for the full clinical evaluation of this new functional imaging methodology in daily routine.
