Improving Valve Representation in Cardiac Stereolithography by Spatially Registering Magnetic Resonance Imaging and Echocardiography
Tyler Robert Moore MDa, Erin Janelle Madriago MDb, and Michael Silberbach MDb
Additive manufacturing, a means of fabricating objects layer by layer though extrusion or sintering, has the potential to impact biomedical research, patient imaging, and medical therapies. One example is personalized anatomic modelling, the creation of tangible models representing the anatomy of an individual patient.1 Of particular interest is the creation of cardiac models in patients with congenital heart malformations.2 The benefits of such models include physician education,3 operative planning,4 and procedure simulation.5
Patient-specific heart models are readily created from cross-sectional data acquired through computed tomography (CT) and magnetic resonance (MR).6 Despite improvements in cardiac-gated CT and MR,7 the temporal resolution of fast moving cardiac structures such as the valve leaflets is limited. Standard three-dimensional echocardiography can acquire volumetric data up to 20 Hz, allowing superior resolution of the valve leaflets.8 Spatially constrained volumetric data is a more recent development in echocardiography that allows for its use in creating heart models through additive manufacturing.9–11 While echocardiography can be used to create very accurate models of the valves, its limited field of view precludes representation of the entire heart, great vessels, and adjacent thoracic structures in a single model.
Integration of multiple modalities allows for more comprehensive modelling of the heart by exploiting both the larger field of view inherent to CT and MR as well as the detailed valve anatomy acquired with echocardiography. Combining modalities requires a means to spatially register the data sets. If there are several anatomic fiducials, corresponding points that are readily identifiable in both data sets, they can be used to determine a linear transformation between the coordinate systems of the two studies.12 Preliminary results have demonstrated the feasibility of combining cardiac MR and three-dimensional echocardiography to create such models.13
Materials and Methods
Subjects are less than 18 years of age. Cardiac models are created from cardiac MR and three-dimensional echocardiography performed in the course of the subjects’ care. The Oregon Health and Science University Institutional Review Board approves maintenance of a cardiac imaging data repository for the creation of heart models for pediatric subjects.
MR is performed using a 1.5 Tesla Philips Ingenia with Philips REV5 software. Sequences used directly for modelling include a pre-gadolinium Fast3D sequence, a pre-gadolinium BFFE cine sequence acquired in the short axis plane that included the atrioventricular valves, and a post-gadolinium angiographic sequence. Gadolinium enhanced sequences are performed using a bolus of Gadavist at 0.1 mmol/kg. Subjects under 12 years of age are routinely sedated by a pediatric anesthesiologist as part of the institutional routine.
Three-dimensional echocardiography data are acquired using a Philips iE33 xMATRIX with an X7-2 probe. Volumetric data sets are acquired in 30 temporal phases per heartbeat with 208 slices per phase. Visual inspection allows retrospective selection of the temporal subset corresponding to the appropriate phase of the cardiac cycle.
Mimics Innovation Suite v17.0 (Materialise, Belgium) is a software package containing tools for segmentation of medical imaging and manipulation of three-dimensional models. Mimics is used for segmenting both MR and echocardiographic data, creation of three-dimensional models from segmented data, spatial registration of models, and conversion of the models to stereolithography format for three-dimensional printing.
Echocardiography data is spatially constrained, but the curvilinear configuration of the probe results in a non-regular grid that is not directly interpretable by the Mimics software. QLAB quantification software (Philips Medical) allows for resampling of echocardiography data from a curvilinear grid into a regular grid, which is exported in a Digital Imaging and Communications in Medicine format. This format contains the spatial metadata that allows three-dimensional analysis in Mimics.
MakerBot Desktop (Makerbot, US) is a software package designed for use with commercially available Makerbot printers. MakerBot Desktop is used to create print instructions for a MakerBot Replicator 2X to convert virtual heart models into tangible objects.
Creating a three-dimensional model from imaging data requires segmentation, the process of selecting the subset of voxels representing the anatomy of interest.
This method uses segmentation techniques established and prescribed by Materialise.2 Mimics contains many tools to expedite segmentation, but several are of particular use in cardiac segmentation. The Threshold tool selects voxels that fall within a user-defined range of values, which quickly segments the blood pool for further refinement. Crop excludes voxels outside user-defined spatial boundaries and Region Grow excludes voxels not directly contiguous with one another. Both are used to quickly excluded portions of the study that are not of interest. The most versatile tool frequently used for segmentation, Multiple Slice Edit, performs operations only on a subset of the voxels. This allows rapid segmentation of low contrast structures such as valve leaflets and septal defects.
The blood pool is segmented from MR data. In a separate Mimics session, the cardiac valves are segmented from echocardiography data. Once completed, the segmentations are converted into surface models called 3D Objects. The 3D Object representing the cardiac valves can be copied into the Mimics session containing the MR-derived blood pool model. The two models are not inherently aligned, however, and must be spatially registered.
Unlike spatial registration involving osseous anatomy, the heart has no high contrast focal structures for use as fiducials. Points along the internal surface of the heart can be used as fiducials, but this requires a technique that accurately identifies identical points along corresponding contours. Points in close proximity to the anatomy being modelled are also preferable because the spatial registration error is lowest near the fiducial centroid.14 The atrioventricular valve annulus meets these requirements, as its contour provides easily identifiable fiducials that are adjacent to the anatomy being modelled.
To determine appropriate fiducials, MR and echocardiography images are first reformatted to a common view that is parallel to the atrioventricular valve plane. The Mimics Online Reslice tool with the Along Plane option performs this task by creating multi-planar reformatted images, which are interpolated images in an orientation different from that of the source data. Viewing the two data sets using the same orientation simplifies identification of corresponding points in the atrioventricular valve plane.
Once fiducials are identifiable in both studies, the Mimics Point Registration tool is used to spatially transform the valve model into the coordinate system of the blood pool model by selecting multiple corresponding points from each data set. This tool performs rigid-body point-based spatial registration using only translational and rotational transformations. The Point Registration tool performs no scaling transformations, which would introduce inappropriate distortion of the appropriately scaled 3D Object.
Unfortunately, the point selection process for the Point Registration tool is limited. This tool only allows selection of points directly from images from one data set. Instead of selecting corresponding points directly from the images of the second data set, points are selected from a three-dimensional object’s surface. Accurately selecting fiducials from the surface of a three-dimensional model is difficult and error prone. To improve the accuracy of fiducial selection from the second data set, a 3D Object representing a single slice through the valve annulus is created alongside the valve model.
The single slice and whole valve 3D Objects are copied from the Mimics session containing echocardiography data into the session containing MR data. Selection of points along this single slice model is much more accurate than selecting points on the surface of the whole valve model because it simulates selection of points directly from a two-dimensional image through the valve annulus.
For older subjects, the spatial and temporal resolution of Fast3D sequence imaging is adequate for both modelling of the heart and determining fiducials for spatial registration.
The faster heart rates of younger subjects requires an approach utilizing two sequences. Pre-gadolinium cine sequences offer excellent resolution of fiducials and are used for spatial registration, but are inadequate for modelling because of the poor spatial resolution that results from their large skip distance. Conversely, post-gadolinium angiographic resolution is too poor for identification of fiducials, but the smaller skip distance allows for adequate heart models. Since younger subjects are routinely sedated and not moved between sequences, a common coordinate system can be assumed between sequences acquired within a study. Spatial registration can therefore be performed using a T1-weighted cine sequence with short axis slices parallel to the AV valve plane, while whole heart modelling can be performed using the post-gadolinium angiographic sequence.
The accuracy of the spatial registration is difficult to evaluate by visual inspection of the 3D Objects. A better qualitative assessment uses the Contours of the valve 3D Object, which displays an overlay of the borders of the 3D Object on the MR images. Accurate spatial registration is confirmed by comparing alignment of structures remote from the fiducials but still represented in both data sets such as the aorta, tendinae chordae, and ventricular walls.
After the blood pool and valves are spatially registered, they are exported into 3-matic, an application in the Mimics Innovation Suite for computer-aided design. This application allows the three-dimensional object to be further refined to better demonstrate anatomical features of interest and to allow for three-dimensional printing. First, the Hollow operation is performed on the blood pool object. This creates a heart model that has an internal surface representative of the internal wall of the heart but with uniform wall thickness. Because the wall thickness is uniform, the resulting heart model does not have an anatomically accurate external surface. However, the uniform wall thickness simplifies three-dimensional printing since plastic extrusion printers fail with thin walled objects. Next, the Cut operation is performed on the valve model using the heart model produced during the previous step as the cutting tool, removing portions of the valve model that are outside the heart model. After this, the models are further refined using the Trim tool to remove the ends of the descending aorta, branches of the aorta, venae cavae, and pulmonary arteries and veins to give them an open-end appearance for easier visual identification as vessels. Finally, Fix Wizard automatically assesses for and corrects problems that can interfere with three-dimensional printing such as overlapping and inverted triangles, mesh defects, and non-manifold meshes. After completion, the models are output as stereolithography files, which are compatible with MakerBot Desktop. Tangible models are then created from acrylonitrile butadiene styrene filament using a Replicator 2X (Makerbot, United States), a commercially available extrusion printer. Since the Replicator 2X has two print heads, models can contain two substrates. The heart model, generated from MR data, is printed in one color while the valve model, generated from echocardiography data, is printed in a second color. The two models are printed simultaneously as a single, fused object.
Results and Discussion
A novel method was employed to create combined echocardiography-MR models for four children aged two, eight, nine, and twelve years. The model from an eight-year-old subject demonstrates an interrupted inferior vena cava, a common atrium, a single ventricle, a single atrioventricular valve, L-malposed great arteries, and pulmonary stenosis (Figure 6). The model from a two-year-old subject demonstrates a large right-sided ventricle that conducts blood to an anteriorly positioned aorta, a ventricular septal defect, a small left-sided ventricle that communicates with a stenotic pulmonary valve, and anomalous pulmonary veins that connect to the right side of a common atrium (Figure 7). This model proved to be particularly valuable in planning this child’s recent successful Fontan operation. The model from a twelve-year-old subject demonstrates normal anatomy with cut planes showing standard echocardiographic views (Figure 8). The model from a nine-year-old subject demonstrates heterotaxy syndrome, L-malposition, pulmonary atresia, and a large ventricular septal defect. This model was of particular clinical utility, changing the management of the patient. Cardiothoracic surgery used the model to consider the relationship between the tricuspid valve and a proposed intracardiac baffle, ultimately deciding that this was not a feasible approach and electing to perform a Fontan operation instead.
While the Mimics Point Registration tool is useful for spatially registering external three-dimensional objects into a study, identifying and selecting appropriate fiducials is challenging. The described approach circumvents this limitation, but a more elegant solution would be modification of the Point Registration tool to allow selection of fiducials directly from images in both data sets rather than from the surface of a three-dimensional object derived from one of the data sets. Additionally, persisting a list of the resulting linear transformations, such as the list of multi-planar reformatted images generated when using the Online Reslice tool, would allow greater flexibility and consistency in spatially registering multiple data sets.
Multiple modalities can be spatially registered using a point-based rigid-body technique to create cardiac models that contain features seen with each modality. When applied to MR and echocardiography data, this technique creates models that demonstrate the entire heart as well as the fast-moving valve leaflets. This approach to personalized cardiac modelling may become increasingly useful. Recently described methods of temporally enhancing three-dimensional echocardiography by frame reordering will potentially improve the resolution of fast moving structures by increasing frame rates to up to 540 Hz. This will increase the feasibility of modelling valves and hearts in even the youngest infants whose heart rates are too fast to image using the current generation of echocardiographic equipment.15
This work was made possible, in part, from a generous grant from The Friends of Doernbecher. The authors wish to thank the applications engineers Todd Pietila, Maureen Schickel, and Eric Renteria at Materialise for help with the Materialise Innovation Suite.
1. Giannatsis J, Dedoussis V. Additive fabrication technologies applied to medicine and health care: a review. Int J Adv Manuf Technol. 2007 Dec 18;40(1-2):116–27.
2. Noecker AM, Chen J-F, Zhou Q, White RD, Kopcak MW, Arruda MJ, et al. Development of Patient-Specific Three-Dimensional Pediatric Cardiac Models: ASAIO J. 2006 May;52(3):349–53.
3. Costello JP, Olivieri LJ, Su L, Krieger A, Alfares F, Thabit O, et al. Incorporating three-dimensional printing into a simulation-based congenital heart disease and critical care training curriculum for resident physicians. Congenit Heart Dis. 2015 Apr;10(2):185–90.
4. Mottl-Link S, Hübler M, Kühne T, Rietdorf U, Krueger JJ, Schnackenburg B, et al. Physical models aiding in complex congenital heart surgery. Ann Thorac Surg. 2008 Jul;86(1):273–7.
5. Shiraishi I, Yamagishi M, Hamaoka K, Fukuzawa M, Yagihara T. Simulative operation on congenital heart disease using rubber-like urethane stereolithographic biomodels based on 3D datasets of multislice computed tomography. Eur J Cardiothorac Surg. 2010 Feb 1;37(2):302–6.
6. Rengier F, Mehndiratta A, Tengg-Kobligk H von, Zechmann CM, Unterhinninghofen R, Kauczor H-U, et al. 3D printing based on imaging data: review of medical applications. Int J Comput Assist Radiol Surg. 2010 May 15;5(4):335–41.
7. Vogel-Claussen J, Pannu H, Spevak PJ, Fishman EK, Bluemke DA. Cardiac Valve Assessment with MR Imaging and 64-Section Multi–Detector Row CT. RadioGraphics. 2006 Nov 1;26(6):1769–84.
8. Hung J, Lang R, Flachskampf F, Shernan SK, McCulloch ML, Adams DB, et al. 3D Echocardiography: A Review of the Current Status and Future Directions. J Am Soc Echocardiogr. 2007 Mar;20(3):213–33.
9. Farooqi KM, Sengupta PP. Echocardiography and three-dimensional printing: sound ideas to touch a heart. J Am Soc Echocardiogr Off Publ Am Soc Echocardiogr. 2015 Apr;28(4):398–403.
10. Mahmood F, Owais K, Montealegre-Gallegos M, Matyal R, Panzica P, Maslow A, et al. Echocardiography derived three-dimensional printing of normal and abnormal mitral annuli. Ann Card Anaesth. 2014 Dec;17(4):279–83.
11. Olivieri LJ, Krieger A, Loke Y-H, Nath DS, Kim PCW, Sable CA. Three-dimensional printing of intracardiac defects from three-dimensional echocardiographic images: feasibility and relative accuracy. J Am Soc Echocardiogr Off Publ Am Soc Echocardiogr. 2015 Apr;28(4):392–7.
12. Boesecke R, Bruckner T, Ende G. Landmark based correlation of medical images. Phys Med Biol. 1990 Jan 1;35(1):121.
13. Moore T, Madriago EJ, Renteria ES, Fuss C, Sahn DJ, Shepard D, et al. Co-registration of 3D echo and MR data to create physical models of congenital heart malformations. J Cardiovasc Magn Reson. 2015 Feb 3;17(Suppl 1):P198.
14. Fitzpatrick JM, West JB, Maurer CR. Predicting error in rigid-body point-based registration. IEEE Trans Med Imaging. 1998 Oct;17(5):694–702.
15. Perrin DP, Vasilyev NV, Marx GR, del Nido PJ. Temporal Enhancement of Three Dimensional Echocardiography by Frame Reordering. Jacc Cardiovasc Imaging. 2012 Mar;5(3):300–4.