Saya Systems logo 3D Data Spatial Registration and
Volume Data Fusion Product
Applied Technology for
Science, Engineering
and Business

Fusion of Volume Data from Multiple Imaging Modalities

Michael P. Zeleznik, Ph.D.

RAHD Oncology Products, St. Louis, MO
Department of Radiation Oncology, University of Utah Medical Center, Salt Lake City, UT

Note: This whitepaper was produced for RAHD Oncology Products and it's commercial version of this tool. However, the tool was jointly developed with:

Marilyn E. Noz, Ph.D. -- Department of Radiology, New York Universitiy Medical Center, New York, NY

Gerald Q. Maguire, Ph.D. - Institute for Microelectronics and Information Technology, Royal Institute of Technology, Kista, Sweden

The Need for Volume Fusion

Advances in the imaging sciences offer new opportunities to improve the diagnosis, staging, and treatment delivery for cancer patients. These new opportunities are dependent on the integration of diagnostic modalities into the treatment planning process, including the need to fuse multiple volumetric image sets into a common coordinate system. Implementation of these technical advances will be impaired if practical considerations are not accounted for. A complete analysis evaluating the merits of a fusion tool should consider the comprehensive use of today's technology; the changes that are certain to occur as technology advances; and the financial implications that are constraints in a very real world. Many fusion tools limit the modalities that may be used, or the tools may only be effective in fusing CT and MR studies of the head. Some fusion tools impose significant constraints for acquiring image data. Obviously, limitations in the fusion tool will decrease its clinical value when the clinician cannot incorporate newly developed imaging techniques into the treatment planning process. The RAHD Volume Fusion tool provides the ability to fuse all modalities without requiring specific acquisition techniques. Cost-per-use will increase if the tools require extensive time in acquiring images or managing the fusion process.

Volume Fusion Basics

Volume fusion is the process of spatially aligning multiple 3-dimensional data sets by transforming them to use a single coordinate system. This process is also referred to as image fusion, or volume/image registration. In general, one set of volumetric data is designated as the reference volume and defines the reference coordinate system to which other volumes will be aligned. The other volumes are then each transformed into alignment with this reference coordinate system. Depending on the data and the application, this transformation can range from simple translation, rotation, or scaling, to the more complex nonlinear warping. The transformation must be based either on externally added information, such as fiducial markers or a stereotactic frame, or on the image data itself. The simplest method of utilizing the image data is to consider the correlation of data between the two volumes. The correlation can be based on a wide range of information, such as the use of known or extracted surfaces, contours, or points; or employing volume techniques based on voxel intensities; or employing statistical methods derived from voxel intensities, such as Mutual Information. Thus, the accuracy of the fusion process depends on 3 issues, which are discussed below:

  1. The method used to correlate the data between the volumes.
  2. The type of transformation employed.
  3. How well the correlation and transformation match the application-specific needs.

Correlation methods fall into two broad categories: automated and manual. Automated registration between modalities which display anatomy clearly, such as computed tomography (CT) and nuclear magnetic resonance imaging (MR), have become commonplace. However, these "automated" methods in other than narrow applications, require operator intervention in almost all cases. Although they generally work very well in the brain, they tend to be less successful in the rest of the body. The situation becomes even more complex when one of the anatomic modalities is replaced with a molecular imaging study (such as PET or SPECT) since the anatomic clues and the information required for successful, automated registration are generally not available. Scintigraphic images have long been used diagnostically to evaluate treatment outcome and recurrence and to aid in planning patient management. New advances in scintigraphic imaging, particularly positron emission tomography (PET) and single photon emission tomography (SPECT), as well as advances in radiopharmaceutical technology are making the use of scintigraphy important even for radiation treatment planning. As external beam placement and brachytherapy loading have become increasingly accurate, the ability to treat the cancer more precisely and limit damage to healthy tissue has led to an increasing interest in using molecular imaging techniques to define active tumor tissue more precisely. Because of the properties of these images, they are not well suited to current automated correlation methods.

The appropriate transformation depends on both the imaging modality and the anatomical region contained in the volumes. For example, fusion of two CT head studies (e.g., pre and post treatment) may require only rotation, translation, and uniform scaling, since typically the shape of the head does not change between scans. But in cases where it does change, for example, following reconstructive surgery, the "automated" methods may need increased operator intervention. Unlike the case of CT head studies, fusion of two CT abdominal studies will require nonlinear warping since the shape of the body, as well as the relative locations of internal structures, may change, even if due only to patient positioning. This becomes even more complex as other modalities are utilized, such as MR, PET, and SPECT. To accurately register with CT data, spatial distortions in MR data can require non-uniform scaling or non-linear warping even for head scans. With SPECT, in the absence of a "transmission" image, there may be no easily visible anatomic structure to register with the CT data. With both SPECT and PET, due to the time taken to complete the scan, patient movement (voluntary or involuntary) can result in a need to warp, especially in areas below the head.

RAHD Volume Fusion

By building upon the research results from NYU Medical Center, RAHD has taken a farsighted approach in creating a new Volume Fusion tool. The primary goal was to provide a tool that is accurate and effective for all imaging modalities and applications, with few constraints on the original volume data. For example, only orthogonal axial slices are required, the patient can be oriented arbitrarily within each volume (e.g., prone and supine), the modalities need not have strong data correlation (e.g., SPECT and CT), and external fiducials are not required. We achieved our goal by employing landmark-based correlation and by supporting a full range of transformations, from simple affine to non-linear warping.

By employing a landmark-based approach, RAHD's Volume Fusion tool successfully fuses most all types of data volumes. This is in stark contrast with other tools that work with only strictly constrained data volumes. Such constraints in other tools include: (1) applicability to only the small subset of situations where automated techniques can work, and further, only to those which do not require warping, (2) requiring external fiducials, or (3) requiring that slices be well-aligned. The first limitation eliminates most non-head studies, and most molecular imaging studies of any area. The latter two are difficult (or costly) to achieve in practice since they require either extra time/effort at the time of image acquisition (or possibly reacquiring the image later) or require additional equipment/materials.

In order to provide accurate fusion of arbitrary volumes, RAHD provides 3 levels of fusion transformation: affine, 1st order warp, and 2nd order warp, in addition to manual affine controls. By providing this powerful flexibility, the operator can maintain fusion accuracy with arbitrary volume data sets. For example, a fusion tool, which only provides rigid body or affine transformation cannot accurately fuse volumes which require 2nd order warping (e.g., abdominal or pelvic studies, or SPECT/PET with CT/MR). Fully automated fusion tools based on mutual information are generally limited to rigid body or affine transformations (especially with molecular imaging studies), as are tools which require manual rotation, translation, and scaling of the volume. Manual volume manipulation is nearly impossible to do effectively on anything except already well-aligned volumes. RAHD recognizes the value of automated techniques such as mutual information (MI) for use in the specific situations where they are known to be effective (e.g., CT/MR head, or CT/CT head). As improvements in computer resources make them practically feasible, these options will be included to augment our tool accordingly.

During development, mathematical simulations of the fusion process were performed on pure mathematical data to ensure the accuracy of the transformation algorithm. Tests were further performed on simulated data sets to ensure that the algorithm behaved as expected on medical data. Extensive clinical tests with real patient data volumes were performed at NYU Medical Center, the Karolinska Institute in Sweden, and the University of Utah Medical Center. These studies include brain, thorax and abdomen volumes. Matches were performed not only between MR-MR, CT-CT and CT-MR but between CT/MR and molecular imaging modalities such as PET and SPECT. Results were evaluated visually, and by measurement of the difference between reference and transformed landmark positions. A third evaluation technique was recently developed. Random sets of 3D points are generated in each volume, and the points falling inside and outside of areas of interest are calculated before and after the transformation, and statistically evaluated.

For the operator, quantitative analysis of the RAHD Fusion results is achieved through error statistics on the difference between reference and transformed landmark positions. Based on this information, the user can quickly find errant landmarks and correct them. Additional analysis of results is possible through (1) precision overlay displays with split views, view ports, or color blending with a variety of colorscales, (2) side-by-side displays with overlayed isolines and landmarks, and (3) simultaneous display of 3D isosurfaces and landmarks.

The RAHD Fusion tool is based on more than a decade of research and development. Our colleagues at NYU Medical Center have been actively researching, using, and refining 2D and 3D warping techniques for over 13 years. Coupled with RAHD's experience in visualization, product development, and customer support, the resulting fusion tool is an accurate and reliable resource with the power to fuse all modalities. Furthermore, RAHD will continue to develop and extend this tool to remain at the vanguard of fusion-based clinical methods for radiation oncology.

Back to Top  
Copyright 2003-2015  Saya Systems Inc. Web design by  Saya Systems Inc.