Statistical coronary motion models for 2D + t/3D registration of X-ray coronary angiography and CTA
Graphical abstract
Highlights
- •
We propose to use coronary motion models for 3D CTA-X-ray angiography registration.
- •
We propose to estimate coronary artery motion from the nearby cardiac surfaces.
- •
This enables building statistical coronary motion models from 4D CTA training sets.
- •
The registration accuracy with statistical motion models and 4D CTA was similar.
- •
The derived motion with statistical models was applicable in a subsequent heart cycle.
Introduction
Percutaneous coronary intervention (PCI) is routinely performed for treating stenosed coronary arteries. During the procedure the site of lumenal narrowing is dilated and a stent is placed under X-ray coronary angiographic (XA) guidance. The XA sequence shows both the guide wire and the vessels due to the injection of contrast agent. PCI on non-occluded lesions has high success rates (99%). These rates drop to 60–80% in the case of treatment of chronic total occlusions (CTOs) (Di Mario et al., 2007, Soon et al., 2007). CTOs cannot be fully visualized with XA as the contrast agent does not pass the occluded segment, which hampers percutaneous treatment. In order to successfully cross a CTO with a guide wire, the interventional cardiologist ideally requires visualization of the occluded segment including the vessel borders and plaque composition, because soft parts of the lesion are easier to cross than calcified parts. Pre-operative CT angiography (CTA) can be used to visualize the entire cardiac vasculature including occluded segments, and the possibility to distinguish calcified plaque (Magro et al., 2010). The combined visualization of the information from CTA and XA during percutaneous treatment of CTOs may increase the success rate of CTO treatment. Therefore, this paper deals with the registration of intra-operative XA sequences with pre-operative 3D CTA images. To account for cardiac motion throughout the registration, we investigate the possibility of employing statistical motion models of the coronary artery.
Many authors have addressed registration of a 3D vascular tree with its projection images. While calibrated biplane acquisitions enable direct reconstruction of the centerline structure (Mourgues, 2001), often only a monoplane setup is available. Registration of the 3D vascular tree with one projection image has been investigated by several authors. Many of the works focus on anatomies which are inherently not moving, such as the cerebral arteries (Feldmar et al., 1995, Kita et al., 1998), or the coronary arteries using an XA frame taken at the same cardiac phase as the CTA image (Metz et al., 2009b, Ruijters et al., 2009, Turgeon et al., 2005). However, selecting the XA frame with the correct cardiac phase is not easy, and an inexact selection may deteriorate registration performance (Turgeon et al., 2005).
To account for slight differences in the vessel shapes, non-rigid 2D–3D alignment techniques have been proposed. Groher et al. (2009) allowed non-rigid deformation of vessel segments of the liver while constraining the length and smoothness of the vessel to ensure plausible deformations. Gatta et al. (2011) corrected for shape differences using 2D non-rigid registration between the projected coronary arteries and one XA frame. Serradell et al. (2011) generated a synthetic motion model of the coronary arteries to constrain the non-rigid motion while fitting the 3D vessel tree to the 2D projection image.
Registration of an entire cardiac cycle rather than a single time-frame provides more information to guide the alignment, but it requires coping with the cardiac and respiratory motion of the coronary arteries. Bouattour et al. (2005) performed frame-by-frame B-spline based 2D–3D fitting, with the previous frame’s fit as the initialization for the next frame. They reported a success rate of 56% on the evaluated three sequences. Metz et al. (2011) proposed simultaneous registration of all frames of a cardiac cycle, using the cardiac motion derived from 4D CTA data. They showed an increase in registration robustness compared to the single frame 2D–3D fitting of the end-diastolic XA frame. However, in clinical practice CTA images are typically acquired within a limited part of the cardiac cycle in order to minimize radiation dose. Therefore, information on the motion of the coronary arteries is often not available. Statistical motion models might in this case replace the patient specific motion.
Statistical motion models have been proposed for respiratory motion tracking in image guided radiotherapy of the liver (Blackall et al., 2001) and the lungs (Liu et al., 2010), in image guided interventions (Klinder et al., 2010), and for modeling heart deformation due to respiration (McLeish et al., 2002). These models were built from multiple images of a single patient to represent the intra-patient variability of breathing. If only one time-point in the cardiac cycle of the patient is known, construction and application of such models is not feasible.
Statistical motion models have also been proposed to represent the motion variation in a population, thus modeling the inter-patient variability. The straightforward extension of statistical shape models to incorporate motion is by concatenating the landmark coordinates of every frame to derive the modeled coordinate vector. Such 4D shape models have been proposed for dynamic segmentation, such as for left ventricle segmentation from ultrasound images (Bosch et al., 2002), and segmentation of the entire heart from 4D CTA (Ordas et al., 2007) and MRI images (Zhang et al., 2010). Others solely modeled motion, thus the difference in point coordinates from a reference frame. Such models have been proposed for recognizing pathologies (Chandrashekara et al., 2003, Perperidis et al., 2005, Suinesiaputra et al., 2009), and for predicting cardiac contraction patterns either for the entire cardiac cycle, or a few phases ahead (Hoogendoorn et al., 2009, Metz et al., 2012).
Coronary artery geometry has previously been studied by Lorenz and von Berg (2006), who built a mean end-diastolic heart surface and coronary artery tree model. Due to the topological variation between patients, the coronary artery model was restricted to the largest arteries, which have a relatively well defined position. A synthetic coronary artery motion model was proposed by Serradell et al. (2011), based on the 3D vessel centerlines. The motion model was generated by hierarchical simulated random movements of the centerline points, where each movement of a point affected the entire sub-tree below that point. The resulting movements remained realistic by smoothing and by constraining movements in the vessel direction. To our knowledge, no population based statistical coronary motion models have been reported in literature. This might be due to the high variability in coronary artery anatomy, and the difficulty to reliably segment coronary arteries in a large number of 4D datasets.
In this paper, we perform 2D + t/3D registration of single-plane XA sequences with pre-operative end-diastolic CTA images. We turn this registration into a 2D + t/3D + t registration by employing a coronary artery motion prior. The primary focus of the paper is the question whether the coronary artery motion prior needs to be patient specific, i.e. derived from a 4D CTA scan (as proposed in Metz et al. (2011)), or could other sources be used, such as a statistical motion model or the mean motion over a population. To investigate this, we propose a methodology to construct patient specific statistical coronary motion models based on 4D CTA images of a training population. The aforementioned problems with population based coronary artery motion models are circumvented by estimating the coronary motion from the motion of the cardiac surface. Four different population based coronary artery motion models are proposed and evaluated for 2D + t/3D alignment: the mean motion, a statistical motion model, the most probable motion based on cardiac shape, and a shape conditional motion model.
The main contributions of this paper are the following:
- 1.
We propose a technique for constructing statistical coronary artery motion models from 4D CTA datasets by coronary motion estimation from the cardiac surface, and generate two types of models: a classic statistical motion model and a shape conditional motion model.
- 2.
We investigate if inclusion of the derived motion models can lead to a similar 2D + t/3D registration accuracy as the 2D + t/3D + t registration employing a 4D CTA based patient specific motion model.
Furthermore, we investigate if the motion estimated with a statistical model can be applied in registration of subsequent cardiac cycles. This is of interest, as a patient specific motion prior recovered from high contrast frames could then be used in subsequent cycles with fading contrast. This paper is an extension of our preliminary work investigating the use of the mean and predicted coronary artery motion estimates for 2D + t/3D + t registration (Baka et al., 2012).
The paper is structured as follows. First, we describe the construction of the proposed statistical coronary motion models (Section 2). Then, in Section 3, we introduce the feature based 2D–3D distance metric used in the registration, and describe the strategy of incorporating different models for the whole cardiac cycle registration. After discussing implementation details in Section 4, and the data in Section 5, the experiments and results are presented in Section 6. Finally, we discuss the results, and present the conclusions in Section 7.
Section snippets
Statistical coronary motion model construction
Building a statistical motion model directly from a training set of 4D coronary artery centerlines would be very challenging due to the variations in coronary topology, which seriously complicates establishing correspondence. Instead, we propose to shift the correspondence definition from the 4D arteries to the 4D cardiac surfaces. We theorize that coronary arteries are attached to the cardiac muscles they are feeding, and therefore move together with that tissue. Coronary artery motion can
2D + t/3D coronary centerline registration
A 2D–3D registration through time consists of two main components: the 2D–3D distance measure evaluating the alignment in one frame, and the strategy in which the frame-by-frame measures are combined and optimized to derive the alignment of the entire cardiac cycle. In this section we describe both components in detail.
Optimization
We used the non-linear least squares optimizer implemented in the optimization toolbox of Matlab R2011. This is a subspace trust-region method based on the interior-reflective Newton method. We set 20 mm and 20° bounds on the initial displacement parameter, and 10 mm and 10 degree bounds on the respiratory model following Shechter et al. (2006). These bounds ensure that the out of plane motion stays within realistic bounds. The motion model parameters were restricted to lie within 3 standard
Training data for the motion model
The 4D heart segmentations from Metz et al. (2012) were used to build the statistical motion models of the coronary arteries. The data consisted of 151 ECG gated 4D CTA images acquired between 2006 and 2010 as part of the clinical diagnosis of patients with acute or stable chest pain, or for research purposes. The data therefore includes a large variety of anatomies and pathologies. 20 of the originally 171 datasets were excluded due to pacemakers (10 cases), and other large anatomical
Experiments and results
We performed two sets of experiments. The first set of experiments focused entirely on the coronary motion model, which was evaluated independently of the task of 2D + t/3D registration. The second set of experiments focused on the 2D + t/3D registration and assessed registration accuracy as well as the quality of the final motion estimate.
Discussion and conclusions
In this paper we investigated the possibility of using coronary motion models combined with a single end-diastolic CTA acquisition for aligning the CTA data with dynamic XA data. Several 4D coronary motion models were investigated, including motion estimates representing a specific motion pattern (such as the mean motion of a population, the predicted motion and no motion), and statistical models, which are able to adapt the motion pattern to the image information.
The proposed statistical
Acknowledgements
This work was financially supported by ITEA Project 09039, Mediate, and NWO Grant Nos. 612.065.618 and 639.022.010. The stitch tracking for removal was performed with Elastix (Klein et al., 2010).
References (37)
- et al.
2D–3D shape reconstruction of the distal femur from stereo X-ray imaging using statistical shape models
Medical Image Analysis
(2011) - et al.
A comprehensive shape model of the heart
Medical Image Analysis
(2006) - et al.
Percutaneous revascularization of chronic total occlusions: review of the role of invasive and non-invasive imaging modalities
International Journal of Cardiology
(2007) - et al.
Comparison of shape regression methods under landmark position uncertainty
- Baka, N., Metz, C., Schultz, C., Neefjes, L., van Geuns, R., Lelieveldt, B., Niessen, W., de Bruijne, M., van Walsum,...
- et al.
A statistical model of respiratory motion and deformation of the liver
- et al.
Automatic segmentation of echocardiographic sequences by active appearance motion models
IEEE Transactions on Medical Imaging
(2002) - et al.
4D reconstruction of coronary arteries from monoplane angiograms
- et al.
Construction of a statistical model for cardiac motion analysis using nonrigid image registration
- et al.
European perspective in the recanalisation of chronic total occlusions (CTO): consensus document from the EuroCTO club
EuroIntervention: Journal of EuroPCR in Collaboration with the Working Group on Interventional Cardiology of the European Society of Cardiology
(2007)
Non-rigid multi-modal registration of coronary arteries using SIFTflow
Deformable 2D–3D registration of vascular structures in a one view scenario
IEEE Transactions on Medical Imaging
The Elements of Statistical Learning (Data Mining, Inference, and Prediction)
Bilinear models for spatio-temporal point distribution analysis
International Journal of Computer Vision
Real-time registration of 3D cerebral vessels to X-ray angiograms
elastix: a toolbox for intensity-based medical image registration
IEEE Transactions on Medical Imaging
Cited by (0)
- 1
These authors contributed equally to this work.