Elsevier

Medical Image Analysis

Volume 8, Issue 3, September 2004, Pages 387-401
Medical Image Analysis

Validation of dynamic heart models obtained using non-linear registration for virtual reality training, planning, and guidance of minimally invasive cardiac surgeries

https://doi.org/10.1016/j.media.2004.06.014Get rights and content

Abstract

Current minimally invasive techniques for beating heart surgery are associated with three major limitations: the shortage of realistic and safe training methods, the process of selecting port locations for optimal target coverage from X-rays and angiograms, and the sole use of the endoscope for instrument navigation in a dynamic and confined 3D environment. To supplement the current surgery training, planning and guidance methods, we continue to develop our Virtual Cardiac Surgery Planning environment (VCSP) – a virtual reality, patient-specific, thoracic cavity model derived from 3D pre-procedural images. In this work, we create and validate dynamic models of the heart and its components. A static model is first generated by segmenting one of the image frames in a given 4D data set. The dynamics of this model are then extracted from the remaining image frames using a non-linear, intensity-based registration algorithm with a choice of six different similarity metrics. The algorithm is validated on an artificial CT image set created using an excised porcine heart, on CT images of canine subjects, and on MR images of human volunteers. We found that with the appropriate choice of similarity metric, our algorithm extracts the motion of the epicardial surface in CT images, or of the myocardium, right atrium, right ventricle, aorta, left atrium, pulmonary arteries, vena cava and epicardial surface in MR images, with a root mean square error in the 1 mm range. These results indicate that our method of modeling the motion of the heart is easily adaptable and sufficiently accurate to meet the requirements for reliable cardiac surgery training, planning, and guidance.

Introduction

Coronary artery disease is the most common cause of death in the developed world. Although not as deadly, atrial fibrillation is also a dangerous condition, accounting for approximately a five to sevenfold increase in stroke risk and a doubling of mortality from cardiovascular disease (Gillinov et al., 2002). For more serious cases of these and similar heart diseases, cardiac surgery remains the only option for the patient. For coronary artery disease, the surgery aims to bypass blockages in the coronary arteries with the aid of grafts, while for atrial fibrillation, the electrical pathways causing the condition are isolated and neutralized by strategic scaring of heart tissue. Access to the heart muscle is therefore required, and is traditionally achieved through a median sternotomy, arresting the heart, and placing it on cardiopulmonary bypass. These highly invasive methods result in additional patient trauma, leading to longer hospital stays and costs compared to less invasive approaches (King et al., 1997). Due to the aging population and the resulting burden on the health care system, it is becoming increasingly important to find treatments for heart disease that are less invasive and more successful than the traditional approaches.

Recent research in endoscopically-aided, port-access, bypass surgeries is showing promising results (Stevens et al., 1996). In these minimally invasive procedures, small incisions (ports) are made in the chest wall to access the heart without requiring a sternotomy. The heart can then be treated manually with long-handled instruments, or with the aid of robotic systems, designed to reduce hand tremor and allow smaller and more appropriate manipulations (Boyd et al., 2000b; Loulmet et al., 2003). Theoretically, this innovative approach to cardiac surgery is a tremendous improvement over the open chest technique, since the surgical target can be treated effectively with less patient trauma. In practice however, minimally invasive cardiac surgery (MICS) is a challenging procedure that requires the surgeon to develop a new and complex set of skills.

Difficulties in maneuvering surgical tools aside, the main challenge in improving the effectiveness of MICS lies in proper visualization of the three-dimensional (3D) operating environment. In spite of the sophistication of the robotic systems used in these procedures, critical tasks such as planning and guidance are often performed entirely with 2D images. The intrinsically 3D port locations are often determined based on 2D angiograms and chest X-rays, while the entire procedure is guided solely through a small field of view video endoscope (Herzog et al., 2003). Even if stereoscopic endoscopy, combined with wide-angle panoramic views are used for navigation, problems can occur due to vision obstruction from water condensation, blood, and tissues. The lack of 3D surgery planning and guidance can lead to improper patient selection, sub-optimal port placement, longer procedures, and increased risks to patients (Boyd et al., 2000a; Herzog et al., 2003).

To address the visual limitations of MICS, we have previously developed a prototype Virtual Cardiac Surgery Planning platform (VCSP) (Chiu et al., 2000). Although currently in the initial stages of development, this project has already yielded promising results. In VCSP, pre-operative computed tomography (CT) and magnetic resonance (MR) images are segmented and stereoscopically rendered in 3D. Models of the patient's ribs and chest wall are created from CT images using the Marching Cubes algorithm (Lorensen and Cline, 1987). A dynamic, 3D, coronary artery model derived from static 3D images and dynamic angiograms (Lehmann et al., 2002) has also been incorporated, and work continues on developing methods to extract patient-specific coronary artery anatomy. Initial models of the patient-specific epicardial surface have been extracted from canine CT images (Wierzbicki and Peters, 2003). We can also overlay the real video endoscopic image with the dynamic virtual environment for the purpose of enlarging the field of view of the endoscope and to allow visualization “through” the tissue (Szpala et al., 2003). Finally, the system can update the position and orientation of virtual surgical tools in VCSP by optically tracking the real instruments. This aspect of VCSP is especially important in surgery planning, where the dependence of inter-costal port locations on the geometry of the simulated surgical tools can be observed and optimized.

One area for VCSP improvement, and the focus of this work, is the development of a method to generate patient-specific, dynamic models of the heart and its components. The addition of such models will make the current prototype VCSP clinically useful for surgery training, planning and navigation. The method must be easily adaptable to different hearts, and be largely automated. The resulting models should also be readily adjustable during a procedure to correct for any changes between the pre- and intra-operative situations. Finally, the resulting models describing the shape and dynamics of the heart must be accurate to within 1 mm root mean square (RMS) error so that it is possible to accurately merge them with models of other structures of interest, such as the relatively small coronary arteries (4 ± 2 mm2 cross-section in women (Kucher et al., 2001)).

Previous heart modeling research has concentrated primarily on the left ventricle because of its importance in the assessment of function (Frangi et al., 2001). A notable exception is the work by Coste-Manière et al. (2003), describing a visualization system similar to VCSP. In their approach, pre-operative CT image segmentation along with 3D reconstruction are used to create static models of the patient's ribs, chest wall and the epicardial surface. These models are then used to automatically calculate the optimal port locations, depending on the target and entry chosen by the surgeon. However, to the best of our knowledge, the significant effect of heart motion is not considered in the current implementation of the epicardial surface model. Automatic segmentation alone can be used to study or display heart dynamics when combined with 4D data acquisition using and ECG-gated protocol. Mitchell et al. (2002) for example, performed frame-by-frame segmentation of cardiac image data to generate models of the left ventricle at different time-points in the cardiac cycle. Although effective on single 3D images, the segmentation approach fails to take advantage of the correspondence of adjacent time-points in a 4D data set. In addition, such methods are usually specific to a particular area of the heart due to their dependence on image intensity. Finally, each image set is usually treated separately so that previous results are not employed to make these methods more adaptable.

Heart parameterization over time and space can be a more adaptable modeling method than the segmentation technique. Essentially, the geometry of the heart (most often the left ventricle) and its evolution over time is assumed to be of a basic shape defined by a series of parameters and equations. Declerck et al. (1998) for example, simplified the geometry of the left ventricle to minimize the number of parameters describing its motion over the cardiac cycle. Such starting models could then be “customized” to fit different patients. These models are suitable for simpler structures such as the left ventricle, but they are difficult to apply accurately in the entire heart, due to the complex motion and geometry.

Assumptions about the shape of the heart are not made when voxel-based, non-linear image registration is used to derive the 3D motion of the heart between two image frames. This motion information can then be used to propagate a static model (created using a single segmentation) through the cardiac cycle to generate a dynamic heart model. An example of such work is given by Lorenzo-Valdés et al. (2002), where the myocardium, left ventricle, and right ventricle are segmented from 4D MR images. Such an approach satisfies the criteria necessary for a dynamic model generation scheme for planning and guidance of MICS. Outside of the creation of the static model, the animation phase is entirely automatic, and the method can be easily adapted to model any part of the heart (depending on the static model used). In addition, the resulting models can be readily adjusted during a procedure with the same type of non-linear registration method used to animate them.

In this manuscript, we present a technique to generate easily adaptable models of the heart using a non-linear registration algorithm. First, a patient-specific, static model of the heart is created by segmenting the end-diastolic (ED) image of a 4D data set acquired with ECG-gating. The model is then animated using motion information obtained by registering the ED image to all remaining time-frames in the 4D data set. The registration is based on a free-form deformation (FFD) framework, where a mesh of deformation points is manipulated to optimize a similarity metric (SM), subject to a smoothness constraint. The main purpose of this work, is a comprehensive validation of the methodology and the resulting model dynamics, in addition to identifying the optimal SM for the registration algorithm. We approach the difficult validation problem with a series of experiments, beginning with registration of a specialized data set obtained from an excised porcine heart. This experiment is important not only because a gold standard is available, but also because it confirms the approach used later to validate the method on in vivo canine and human data sets.

Section snippets

Cardiac models

In the following subsections we discuss how a dynamic surface model of the heart epicardium or endocardium is created. The process begins with an ECG-gated, 4D image acquisition. The resulting data set consists of several 3D images, each depicting the heart at different time-points in the cardiac cycle. The motion information is extracted from the 4D data using a non-linear registration algorithm described in detail in Section 3. The static geometry of the heart is obtained from one of the 3D

Non-linear registration algorithm

This section describes the registration algorithm used to extract the motion of the heart between two image frames of a 4D data set. This description is broken into several subsections, beginning with a discussion of the non-linear transformation model, then a description of the solution method, and finally a description of the fundamental image properties and mathematics used in obtaining the solution.

Validation methods

Validation of non-linear registration methods is difficult since a gold standard is not easily attainable. Researchers are forced to use a variety of surrogates, from simplified computer simulated images, to difficult experiments involving implantation of fiducials into the subject (Woods et al., 1998a, Woods et al., 1998b). The heart poses additional validation problems due to a lack of easily identifiably homologous points (Mäkelä et al., 2002). Yet validation is arguably the most important

Results

Original and down-sampled images of the excised porcine heart were registered using the approach discussed above. All registrations were repeated with each of the SMs introduced in Section 3.3. The error is expressed as the RMS of the Euclidian distance between the landmark positions in the source image after registration, and the landmark positions in the target image. The results are presented in Table 1.

For the canine data sets we used the consistency measure to estimate error in extracted

Discussion

In this manuscript, we extracted motion from 4D heart images using a registration algorithm with several different SMs. The accuracy of the resulting motion was then validated yielding several interesting results that are worth noting. First, the accuracy of the registration is significantly better if the input images are initially more similar. For example, registration 5 to 0 in Table 1 is more challenging than registration 5 to 4. Similar patterns were observed in the in vivo experiments

Conclusions and future work

In this work, dynamic models of the heart were obtained from CT and MR images, through a combination of segmentation, surface rendering, and image registration. The accuracy of the cardiac models was determined with three validation experiments on ex vivo, and in vivo canine and human data sets. On average, the models satisfied our accuracy requirement of less than 1 mm error when the correct SM was chosen. In the excised heart experiment we showed that MI gives the best accuracy, probably due

Acknowledgements

The authors thank Aaron So, Rhonda Walcarius, and Kathy Parker for help image acquisition, Joy Dunmore-Buyze for phantom preparation, Atamai, Inc. for visualization software, Ravi Gupta for help with code development, Dr. Mark Wachowiak, John Moore and Guy-Anne Turgeon for helpful discussions and assistance in writing this manuscript. This work was supported by grants from the Canadian Institutes of Health Research (MOP 14735), Canadian Heart and Stroke Foundation (NA 4755), Ontario Consortium

References (32)

  • F.L. Bookstein

    Principal warps: thin-plate splines and the decomposition of deformations

    IEEE Trans. Patt. Anal. Mach. Int.

    (1989)
  • W.D. Boyd et al.

    A comparison of robot-assisted versus manually constructed endoscopic coronary anastomosis

    Ann. Thorac. Surg.

    (2000)
  • A.M. Chiu et al.

    3-D image guidance for minimally invasive robotic coronary artery bypass

    Heart Surg. Forum

    (2000)
  • A.F. Frangi et al.

    Three-dimensional modeling of functional analysis of cardiac images: a review

    IEEE Trans. Med. Imaging

    (2001)
  • C. Herzog et al.

    Multi-detector row CT versus coronary angiography: preoperative evaluation before totally endoscopic coronary artery bypass grafting

    Radiology

    (2003)
  • D.L.G. Hill et al.

    Medical image registration

    Phys. Med. Biol.

    (2001)
  • Cited by (95)

    • Characterization of the Physiological Displacement of the Aortic Arch Using Non-Rigid Registration and MR Imaging

      2017, European Journal of Vascular and Endovascular Surgery
      Citation Excerpt :

      In general, overall motion is studied by an affine transformation and localized deformations are assessed by elastic transformations based on b-spline models. The main applications of these methods are motion compensation in interventional cardiology surgery6–8 and prediction of deformations during endovascular abdominal aneurysm aortic repair.9 Faranesh et al. used affine registration to calculate motion fields to deform a reference end-diastolic image to different cardiac phases.6

    • On mixed reality environments for minimally invasive therapy guidance: Systems architecture, successes and challenges in their implementation from laboratory to clinic

      2013, Computerized Medical Imaging and Graphics
      Citation Excerpt :

      Nevertheless, it has been suggested that such technology would have a major impact on some of the most life-threatening conditions that require intervention [122]. Several examples of procedures that incorporate non-rigid registration include the use of statistical atlases for neurosurgical interventions [123]; deformable atlases of bony anatomy for spine interventions [124,125]; motion models for image-guided radio-therapy of the lung [126] and liver [127]; and interventional cardiac applications [128,92,129]. Despite the wealth of available registration algorithms, some approaches may not be suitable for use in time-critical interventional applications in the OR.

    View all citing articles on Scopus
    View full text