Automated detection of intracranial aneurysms based on parent vessel 3D analysis
Introduction
Subarachnoid hemorrhage (SAH) is a serious cause of stroke which affects 30,000 patients in North America annually. SAH accounts for a quarter of cerebrovascular deaths, with 80% of the non-traumatic SAH cases being caused by a ruptured intracranial aneurysm (Edlow et al., 2008, Wardlaw and White, 2000). An intracranial aneurysm is a localized pathological dilatation of a blood vessel. It is reported that up to 2% of the general population harbors aneurysms (Rinkel et al., 1998, Juvela, 2004). Most of these aneurysms are asymptomatic and remain undetected with only a small proportion proceeding to rupture and consequent SAH, with an annual incidence of approximately 1% (Wardlaw and White, 2000, Johnson et al., 2001). However, in the case of a ruptured aneurysm, the initial bleed is fatal in 10–20% of instances and despite improvements in patient management, the incidence of SAH has not declined over time and the morbidity rate is still reported between 25% and 50% in patients surviving aneurysm ruptures (Wardlaw and White, 2000, Juvela, 2004, Suarez et al., 2006).
Detecting intracranial aneurysms from imaging scans is an essential step in the prevention of aneurysmal SAH and its attendant complications (Wardlaw and White, 2000), as treatment of aneurysms using endovascular or surgical methods carries a lower rate of complication when performed in unruptured versus ruptured aneurysms (Brisman et al., 2006). Although aneurysm detection is currently performed visually by experienced diagnosticians, there is an increasing interest in computed-aided diagnostic (CAD) systems to assist diagnosticians and possibly improve diagnostic accuracy, while limiting missed detection.
The purpose of this work is the introduction and initial proof of concept of a new 3D shape feature, namely the writhe number, used for detecting aneurysms in high quality segmentation of vasculatures. As such our contribution is twofold. First, we concentrate on our theoretical contribution by introducing the writhe number as a new 3D descriptor used to characterize surfaces. Known in curve theory since its introduction by Fuller (1971), the writhe number is used to describe the global geometry of a closed space curve or knot (Agarwal et al., 2004, Berger and Prior, 2006). To the best of our knowledge, this paper represents the first time the writhe number is extended to surfaces. Second, we develop a writhe number-based scheme for the automatic detection of aneurysms and demonstrate its utility via the analysis of both 3D-RA and CTA data.
Existing aneurysm detection methods focus on magnetic resonance angiography (MRA) data and are usually two-step processes (Arimura et al., 2004, Uchiyama et al., 2005, Kobashi et al., 2006). First, potential regions of interest (potential aneurysms) are detected by pre-processing the data using dot-enhancement filters (Arimura et al., 2004) and/or by analyzing the geometry of the vessels. Second, false positive reduction methods are applied on the areas highlighted in the first step, where the reduction scheme depends on the specificity of the detection method used in the first step.
The detection method presented here is based on the 3D shape analysis of the vessels and it is performed on the segmented vasculature. We assume that a short segment of a normal vessel can be locally modeled as a tube with a circular cross-section whose medial axis is a line segment or a quadratic (i.e., parabolic) curve. This assumption is validated experimentally on clinical data. Using the writhe number our method identifies aneurysms as regions of the vasculature which are not well modeled as a tube or an extruded parabola. Specifically, regions along the vessels where the writhe number is non-zero are reported as possible aneurysms. The method uses a false positive reduction scheme in which small regions are eliminated from positive results.
In this study, the detection method is tested on 3D-RA and CTA patient data. The work concentrates on the use of a high resolution modality, in this case 3D-RA, to enable the evaluation of the algorithm characteristics without being hampered with the inevitably lower spatial resolution of most current MRA and CTA data. We also consider the application of the algorithm to CTA data and discuss its performance and limitations on this more challenging imaging modality.
Free-response operator characteristic (FROC) analysis is applied to evaluate the performance of the proposed detection system. FROC analysis shows how the sensitivity of the system changes function of the threshold value used to eliminate small positive results. In our experiments, the sensitivity of the aneurysm detection method was found to be 100% with 0.66 false positives per study on ten distinct 3D-RA datasets and 5.36 false positives per study on ten unrelated CTA datasets.
The eventual clinical goal of this research is to offer an added safety net to the diagnostician and the patient, by making available a concordance check protocol that would point the clinician to potential areas of concern that may have been missed by the current method of visual inspection. The added value of such a tool will need to be evaluated by prospective clinical trials.
The paper is structured as follows: existing work in vessel segmentation and aneurysm detection is presented in Section 2. In Section 3 we introduce the writhe number followed by details about the detection method in Section 4. Test data and pre-processing procedures are presented in Section 5. Results are reported in Section 6 and discussed in Section 7, together with directions for future work. Proofs involving the writhe number are demonstrated in Appendix A.
Section snippets
Related work
When interpreting scans and searching for aneurysms, it is important for clinicians to have access to the underlying 3D structures from the 2D studies. Because 3D-RA, CTA and MRA data provide vessel and aneurysm positions in cross-sectional images only, the extraction of 3D structures from 2D images is achieved through segmentation. A great deal of research has been carried out in developing algorithms for the segmentation of cerebral vasculature, including aneurysms, from MRA and CTA studies (
The writhe of surfaces
The writhe number was introduced by Fuller (1971) and it is used in curve theory to measure how much a curve twists and coils. When a second curve is placed nearly parallel to the first one, the writhe number measures how much the second curve twists about the first (Berger and Prior, 2006). In biomedical engineering, the writhe number is used to study the shape and topology of DNA (Klenin and Langowski, 2000, Rossetto and Maggs, 2003) or to characterize the shape of curves on 3D surfaces, such
Overview of the detection algorithm
The detection method takes as input a 3D volume in which the cerebral vasculature has been segmented from the background. The medial axis of the vessels is computed from the segmented volume. Similar to Arimura et al. (2005), we consider that aneurysms appear as short branches in the medial axis of the vasculature. Local neighborhoods are determined for surface points along short branches such that they satisfy the connectedness and size conditions described in Section 3.1. The writhe numbers
3D-RA and CTA
The aneurysm detection method was tested on ten distinct 3D-RA and ten unrelated CTA patient-derived datasets. The twenty datasets contain twenty aneurysms, with one study showing no aneurysms and one study having two aneurysms. The aneurysms have diameters in the range 3.2–10.2 mm and lengths in the range 3.5–13 mm. Among the aneurysms, six are sidewall aneurysms (dilation of the artery in one direction perpendicular to the vessel axis), nine are bifurcation aneurysms (dilation at the
Results
All aneurysms were correctly identified by our detection method with 0.66 false positives per study on 3D-RA data and 5.36 false positives per study on CTA-derived data. These results were obtained as follows. As discussed in Section 4.7, we start by clustering voxels whose writhe number is non-zero and then computing the region index associated with each cluster. Suspect regions are taken as those whose region index exceeds a given threshold. The performance analysis in this paper is evaluated
Discussions
The proposed aneurysm detection method was tested on twenty datasets from two imaging modalities, acquired with three different scanner models.
As shown by the FROC analysis, the detection algorithm performs very well on 3D-RA data and results in few false positive results (0.66 per study). 3D-RA images have high resolution and show high contrast between vasculature and surrounding tissue and simple segmentation techniques result in accurate segmented volumes. Segmentation is more challenging on
Conclusion
A new method is presented for automated detection of intracranial aneurysms, which is based on the local 3D shape of the parent vessels. The writhe number is introduced as a new surface descriptor that can be used to distinguish between tubular and non-tubular regions along the vessels. The detection algorithm requires only a segmented volume of cerebral vasculature and is otherwise independent of the imaging modality. The method is tested on 3D-RA and CTA patient data. The robustness of the
References (29)
- et al.
Automated computerized scheme for detection of unruptured intracranial aneurysms in three-dimensional magnetic resonance angiography
Academic Radiology
(2004) - et al.
Flux driven automatic centerline extraction
Medical Image Analysis
(2005) - et al.
Aneurysmal subarachnoid hemorrhage: update for emergency physicians
Journal of Emergency Medicine
(2008) - et al.
Extracting branching tubular object geometry via cores
Medical Image Analysis
(2004) - et al.
Non-parametric geodesic active regions: method and evaluation for cerebral aneurysms segmentation in 3dra and cta
Medical Image Analysis
(2007) - et al.
Computing the writhing number of a polygonal knot
Discrete and Computational Geometry
(2004) Least squares orthogonal distance fitting of curves and surfaces in space
(2004)- et al.
CAD scheme for detection of intracranial aneurysms in MRA based on 3D analysis of vessels skeletons and enhanced aneurysms
Proceedings of SPIE
(2005) - et al.
Computerized detection of intracranial aneurysms for three-dimensional MR angiography: Feature extraction of small protrusions based on a shape-based difference image technique
Medical Physics
(2006) - et al.
The writhe of open and closed curves
Journal of Physics A: Mathematical and General
(2006)