Elsevier

Medical Image Analysis

Volume 18, Issue 1, January 2014, Pages 83-102
Medical Image Analysis

Semi-automatic segmentation and detection of aorta dissection wall in MDCT angiography,☆☆

https://doi.org/10.1016/j.media.2013.09.004Get rights and content

Highlights

  • A method for semi-automatic segmentation of the aortic dissection wall is proposed.

  • The centerlines of the aorta and its main branches are extracted semi-automatically.

  • We segment the outer vessel wall using a geodesic levelset framework.

  • An algorithm is proposed to extract the dissection wall as a 3D mesh.

  • Results on five MDCT datasets show an mean absolute error of 0.34 mm.

Abstract

Aorta dissection is a serious vascular disease produced by a rupture of the tunica intima of the vessel wall that can be lethal to the patient. The related diagnosis is strongly based on images, where the multi-detector CT is the most generally used modality. We aim at developing a semi-automatic segmentation tool for aorta dissections, which will isolate the dissection (or flap) from the rest of the vascular structure. The proposed method is based on different stages, the first one being the semi-automatic extraction of the aorta centerline and its main branches, allowing an subsequent automatic segmentation of the outer wall of the aorta, based on a geodesic level set framework. This segmentation is then followed by an extraction the center of the dissected wall as a 3D mesh using an original algorithm based on the zero crossing of two vector fields. Our method has been applied to five datasets from three patients with chronic aortic dissection. The comparison with manually segmented dissections shows an average absolute distance value of about half a voxel. We believe that the proposed method, which tries to solve a problem that has attracted little attention to the medical image processing community, provides a new and interesting tool to isolate the intimal flap that can provide very useful information to the clinician.

Introduction

The aorta is the thickest artery of the body and the one that delivers most blood. The gradual increase of life expectancy in developed countries has led to an increased occurrence of aorta diseases, including aneurysms, dissections and coarctation among others. Thoracic aorta diseases can be asymptomatic like aneurysms, which are often observed accidentally while looking for other diseases, or present clear symptoms with pain like dissections. The dissection of the thoracic and abdominal aorta is a serious vascular disease produced by the rupture of the tunica intima, which is the innermost layer of the artery wall, and can be lethal to the patient.

The rupture of the internal layer can let the blood flow between the different wall layers and divide the vessels in two parts, creating two lumens called true and false lumens. The true lumen refers to the one that starts from the left ventricle and leads the blood to the inferior members, while the false lumen is a cul-de-sac that just retains the blood but does not bring it anywhere. This process can lead to a pressure on the true lumen that could prevent the blood from flowing neither through the true lumen nor through its ramifications. In some cases, visceral arteries can be fed by the false lumen.

The aortic arch extends from the proximal part of the origin of the ascending aorta to the distal part of the origin of the left subclavian artery. It is the most important anatomical area for the classification of the dissections into type A or B depending on their occurrence before or after the supra-aortic trunk. The anatomical classification is based on the involved aorta segment. The strength of the anatomical classification depends on its ability to determine the prognostic and to drive the patient treatment in a individual way. The Stanford classification (Daily et al., 1970) is currently the main one, and resumes and simplifies the previous classification of DeBakey et al. (1965), by considering two types of dissections, A and B (as shown in Fig. 1):

  • Type A involves the ascending aorta and corresponds to types I and II of DeBakey classification,

  • Type B does not involve the ascending aorta and the lesion is localized after the left subclavian branch. It corresponds to types III and IV of DeBakey classification.

An aortic dissection can also bring an associated swelling or aneurysm of the aorta. The exact reason is unknown, but it is likely due to atherosclerosis o arterial hypertension. Traumatic lesions are the main causes of aortic dissections. The decision and choice of treatment related to aorta dissections is currently a controversial issue, especially regarding the peripheral complications of the malperfusion of the digestive organs, of the kidney and of the inferior members (Deeb et al., 1997, Laas et al., 1991). These complications are the origin of a mortality of 20% of type A dissections and of 10% of type B ones. The apparition of endovascular treatments has changed the decision tree related to these diseases. As a consequence, the traditional dogma of surgical operations for type A dissections and medical treatment for type B one has changed. Computed Tomography (CT) enables a fast evaluation of the patient and is an important tool to diagnose patients with thoracic pain. The vascular screening with angio-CT includes the injection of intravenous contrast agent that brings a high contrast between the vascular structures and the rest of the image. Multi-detector CT or MDCT have considerably improved the image resolution in the Z axis (Catalano et al., 2003, Rubin, 2003, Kapoor et al., 2004). Together with the speed in the administration of the contrast, in the motion of the table which holds the patient, and in the image acquisition, MDCT becomes a key modality to study the aorta, allowing the acquisition of high resolution and high quality images with a minimal effort from the patient, which is asked to make a short apnea to improve the study.

In this work, we propose a semi-automatic method for segmenting thoracic and abdominal aorta dissections from angio-MDCT images.

We can distinguish two main tasks required for the analysis of three-dimensional vascular structures, and the correct visualization and quantification of the images: the extraction of vessels centerlines, which define the vascular topology, and the segmentation of vascular structures, which usually consists in the extraction of the lumen. Several strategies can be used to obtain both the centerlines and the segmented vessels:

  • start with an extraction of the centerlines, for example using minimal cost path as in Deschamps and Cohen (2001), and then use those centerlines as a basis for the segmentation (Krissian and Arencibia, 2009, Schaap et al., 2009b);

  • first obtain a segmentation of the vascular structures using a segmentation algorithm that can be geometric (Sethian, 1999, Krissian and Westin, 2005) or statistics (Chung et al., 2004) and define the centerlines from this segmentation;

  • obtain both the centerlines and the segmentation at the same time using tracking algorithms (Friman et al., 2010, Florin et al., 2005).

One image processing task that is often used in any of these strategies is the enhancement of vascular structures based on a filter that can use the Hessian matrix to characterize tubular structures and that is usually applied within a multiscale analysis process to combine responses from structures of different sizes or radii. The image obtained from this multiscale analysis can be used as an input for other algorithms that extract the vessels centerlines automatically or semi-automatically. One of these techniques consists in looking for the minimal cost path between two spatial locations manually selected by the user.

A more detailed state-of-the-art in vascular segmentation can be found in (Kirbas and Quek, 2004, Lesage et al., 2009). Although most techniques achieve good results in images of good quality, it is in general very difficult to get a full and accurate extraction of the whole vascular network for the following reasons:

  • Vascular structures, although mainly cylindrical with circular cross-sections, can have a wide range of different shapes and sizes, curvatures and tortuosities, and form a tree-like structure including bifurcations, usually resulting in structures that span over most of the image and are represented in a wide range of intensities. They also extend to most parts of the human body, and can be interwoven (arterial and venous system), and get closer to other structures that have a similar or same intensity (bones, heart, …).

  • Each acquisition modality brings its own artifacts like noise, spatial distortions, and intensity changes due to interferences with other structures of the human body.

  • In pathological areas like stenosis, aneurysms or calcifications, vessels can change shape rapidly, which harden their automatic detection.

  • Part of the vessel can be coagulated or obstructed, preventing the blood from flowing (thrombosis).

In the last years, several initiatives have been presented to compare existing vascular segmentation methods in the form of challenges where different methods can compete on the same input datasets for a specific segmentation task. One of them consisted in the extraction of the centerlines from coronary arteries (Schaap et al., 2009a), another one focused on the segmentation of the carotid artery bifurcation (Hameeteman et al., 2011). Many techniques have also been proposed in the context of vascular segmentation in CT angiography, applied to different vascular trees like the coronary arteries (Metz et al., 2009) and the aorta (Saur et al., 2008).

MDCT is usually considered as the image modality of choice to acquire high definition images of a dissected aorta. In practice, the patient diagnosis is often complex (Willoteaux et al., 2004, Gaxotte et al., 2003) and is performed directly on the 2D orthogonal slice sections, combined with Volume Rendering, Multi-Planar Reconstruction (MPR), and Virtual Endoscopy. However, none of these techniques is able to depict the dissection wall as a whole, since it appears as a thin black structure within the contrast-enhanced vessel lumen. For this reason, an automatic segmentation and 3D visualization of the dissection wall could provide the physician with key new tools that improve the quality of the diagnosis while reducing its time consumption.

Nevertheless, this topic has received little attention from the medical image processing community. Among the previous works related to the detection and segmentation of the dissection wall, authors have focused on computing displacement of the intra-arterial septum (or internal flap) from in vivo 2D phase-contrast magnetic resonance sequences (Karmonik et al., 2009), extracting true and false lumen based on multiscale wavelet analysis and generative-discriminative model matching (Lee et al., 2008), extracting the dissection wall contours in 2D with a subpixel accuracy (Trujillo-Pino et al., 2012), and extracting the dissection wall based on a multiscale sheetness measure and an automatic segmentation of the aorta (Kovács et al., 2006b, Kovács et al., 2006a).

We propose to extract the dissection wall by combining modified versions of existing algorithms to achieve several preprocessing tasks like noise reduction, extraction of the vessels centerlines and segmentation of the outer wall of vascular structures based on the centerlines.

Once the outer wall of the aorta and its main branches have been extracted, we introduce a new algorithm to extract the center of the dissection wall as a polygonal mesh inside the segmented vessels, defined as a zero-crossing of the scalar product between two vector fields, after re-orientation of one of them to ensure its spatial continuity. After manual segmentation of the dissection wall in our five datasets, we show that the result of our algorithm is very close to the expert segmented walls, with a global average absolute distance of about half a voxel.

Our study has followed the local ethical rules and the ethical norms from the World Medical Association. Images have been acquired in the clinical hospital with a multi-detector CT model LightSpeed VCT from GE Medical Systems, with 64-row detectors (64 rows × 888 detector elements). Our database consists of 6 datasets, from 4 different patients, of which 2 are women and 2 are men with an average age of 63 years, as shown in Table 1 and Fig. 2. The images from the first patient (001) are post-operative acquisitions from a type A dissection, where the patient has undergone a replacement of the ascending aorta with a Dacron tubular graft.

The last dataset consists of a normal patient without dissection, it will be used as an example to demonstrate the possible false positives obtained by the proposed method. The other patients present chronicle dissections and have been selected consecutively. However, we did not consider patients with a thrombosed dissection, where the intensity of the false lumen drops considerably and hardens the automatic detection of the outer aortic wall. All the patients have been scanned through angio-MDCT after injection of intravenous contrast and with 0.6 mm slices. The datasets have been acquired in DICOM format, with a slice resolution of 512 × 512 pixels, and an energy range from 100 kVp and 120 kVp depending on the patient weight.

All the patients have been inspected by an expert in vascular radiology who isolated the intimal flap in all the cases.

Section snippets

Method

We propose to use a combination of existing and new techniques to allow a semi-automatic, robust and efficient delineation of the internal dissected aorta wall. The general scheme of the proposed method is illustrated in Fig. 3.

Our method for extracting and visualizing the dissection wall is divided in two main steps: the first one consists in preprocessing the initial dataset and extracting the centerlines of the aorta and its main branches; in the second step, we segment the outer wall of the

Synthetic images

To evaluate the quality of our method, we created synthetic images representing a dissected vessel in different configurations. The synthetic images have about 1003 voxels and simulate an aorta lumen of intensity 300 and dissection and background of intensity 80, corrupted with additive Gaussian noise of standard deviation 35. The image have been generated with a partial volume effect and intensities similar to the intensities present in the real computed tomography datasets (see Fig. 13).

The

Conclusion

Very little work has been dedicated to the extraction of the aorta dissection wall, which is a serious disease that is usually complicated for the physician to interpret, since no current three-dimensional view of the dissected wall is available. In this work, we propose a semi-automatic method to extract the dissection wall as a 3D polygonal mesh, allowing its 3D visualization that can be combined with volume or surface rendering of the data or the segmented aorta with transparency. As a

References (47)

  • M. Schaap et al.

    Standardized evaluation methodology and reference database for evaluating coronary artery centerline extraction algorithms

    Medical Image Analysis

    (2009)
  • Y. Boykov et al.

    Fast approximate energy minimization via graph cuts

    IEEE Transactions on Pattern Analysis and Machine Intelligence

    (2001)
  • Buades, A., Coll, B., Morel, J.M., 2005. A non-local algorithm for image denoising. In: IEEE Computer Society...
  • V. Caselles et al.

    Geodesic active contours

    International Journal of Computer Vision

    (1997)
  • C. Catalano et al.

    MDCT of the abdominal aorta: basics, technical improvements, and clinical applications

    European Radiology

    (2003)
  • A. Chung et al.

    Vascular segmentation of phase contrast magnetic resonance angiograms based on statistical mixture modeling and local phase coherence

    IEEE Transactions on Medical Imaging

    (2004)
  • T. Deschamps et al.

    Fast extraction of minimal paths in 3D images and application to virtual endoscopy

    Medical Image Analysis

    (2001)
  • Farnebäck, G., 2002. Polynomial Expansion for Orientation and Motion Estimation. Ph.D. thesis, Linköpings unicersitet....
  • C. Florin et al.

    Particle filters, a quasi-monte carlo solution for segmentation of coronaries

  • V. Gaxotte et al.

    Relationship of intimal flap position to endovascular treatment of malperfusion syndromes in aortic dissection

    Journal of Endovascular Therapy

    (2003)
  • V. Kapoor et al.

    Intimomedial rupture: a new CT finding to distinguish true from false lumen in aortic dissection

    American Journal of Roentgenology

    (2004)
  • Karmonik, C., Bismuth, J., Davies, M., Younes, H., Lumsden, A., 2009. An image processing algorithm for the in-vivo...
  • M. Kass et al.

    Snakes: active contour models

    International Journal of Computer Vision

    (1988)
  • Cited by (0)

    This is a collaborative effort.

    ☆☆

    This work has been supported by the project TIN2009-10770 (Spanish Ministry of Science and Innovation), the Ramón y Cajal program from the Spanish government, and the Motiva project from the government of the Canary Islands.

    View full text