Virtual breast biopsy

We are developing virtual breast biopsy, a non-invasive method of probing a suspicious breast lesion without surgery, thereby avoiding pain, bruising, and scarring.  Most breast biopsies are performed on what turn out to be benign breast lesions. Instead of surgical biopsy, we are using a combination of Magnetic Resonance Mammography (MRM)
      http://my.webmd.com/hw/health_guide_atoz/stm159519.asp?navbar=aa80345
      http://www.emedicine.com/radio/topic792.htm
and Positron Emission Tomography (PET)
      http://my.webmd.com/hw/health_guide_atoz/aa80345.asp
with F-18-fluoro-deoxyglucose (FDG). Our research has provided strong indications that the diagnostic value of combined MR and PET mammography, which we call virtual breast biopsy, could be as high as that of real surgical breast biopsy. We are investigating what are the most suitable MRM and PET tests, and what are the best ways of combining and using them as virtual breast biopsy.
 

This research includes three main projects:

Nonrigid PET/MRI Breast Image Registration and Fusion
There are many different types of images that can be obtained from MRM and PET. The diagnostic value of combined MR and PET mammography depends on the selection of the images and on the way the images are fused together. For virtual breast biopsy, we need to select the most suitable MRM and PET images, precisely align (register http://imaging.birjournals.org/cgi/content/figsonly/14/6/455 ) them in three-dimensions (3D) and then, display them to physicians, using a system for intelligent display of the combined images. Such an intelligent display system would allow easy discovery of features indicating a possible malignant lesion and would show how much of this information is provided by different imaging modalities. For example a region where an apparently malignant lesion appears in PET might appear unremarkable in MRM, and it is very important that a physician be given this information.
The MRM and PET breast images are obtained with the patient prone and breasts gently placed within two open “cups” for support (Fig. 1). In MRM studies these cups house a special RF receiver coil needed to produce the images. 
 

Fig. 1.
Fig. 1. Example of a dedicated 4-element phase array breast antenna, compatible with Siemens iPAT (parallel receiver technology).
 

In PET studies, the cups are needed to assure that the breast tissue is imaged in the same configuration as in MRM. Specially prepared fiducial skin markers are taped in predefined locations on the breasts. These markers are visible in both MRM and PET images, and are needed for accurate fusion of the images (Fig. 2). 
 

Fig. 2.
Fig. 2. Coronal cross-section through MRM and PET breast images. Green: MRM. Red: PET. Arrows mark the difference (displacement) vectors between corresponding fiducial skin markers visible in MRM and PET.
 

Because the breasts are entirely composed of soft tissue, they easily deform and require nonrigid image registration, for which we have developed an iterative approach. In this process, a PET scan provides the target image (remains unchanged) while MRI provides an image that is warped to match the target image. This process consists of two iterations (Fig. 3). 
 

Fig. 3.
Fig. 3. Block diagram of our iterative FEM deformable model for nonrigid registration and fusion of PET and MRM breast images with the aid of fiducial skin markers.     (click on block diagram to see full size)
 

In the first iteration, corresponding fiducial skin markers (FSMs) are identified on both PET and MRM images, and locations of their centroids are estimated using a knowledge based semi-automated algorithm. This allows estimation of FSM displacement vectors between PET and MRI (Fig. 2). 

Patient-specific geometry of the breast is obtained from sagittal or coronal cross sections of the MRM images. In order to obtained warped images, the 3D image of breast tissue is converted to a mesh (special grid) by use of a finite element method (FEM), using triangular surface elements and tetrahedral volume elements (Fig. 4)
 
Fig. 4a.
a   (click on image to see larger view)
Fig. 4b.
b   (click on image to see larger view)
Fig. 4. Example of a patient-specific breast tissue 3D image converted to a FEM mesh. The original 3D breast image was obtained using MRM.

A deformed FEM mesh is then obtained using FEM, by first distributing linearly the Cartesian components (separately along X, Y, and Z axes) of FSM displacement vectors over the breast surface, and then throughout the volume. Our FEM model uses an analogy between geometrical mesh deformation and the temperature distribution in a solid object. A warped MRI breast image is then obtained using the deformed mesh. It is now registered in 3D to the target PET breast image. However, there are still small inaccuracies in the regions away from FSMs (Fig. 5).
 


Fig. 5. Projections of error of surface of breast matching between MRM and PET on an arbitrary plane approximately parallel to the chest wall. Left panel: After rigid registration. Center panel: After first iteration of registration. Right panel: After second iteration of registration (surface refinement). White dots indicate FSM locations. The error in surface registration is encoded using a color scale (see calibration bar on the right). Black indicates no error; red indicates 4 mm error in surface registration.
 

To address this problem, we perform a second iteration, surface refinement, in which a large number of corresponding surface points are determined on the warped MRM image and the CT image of the breast that is obtained just before PET (using a PET/CT scanner) and that is coregistered with the PET images. This is done because the breast surface is poorly defined in PET studies. The displacement vectors between the corresponding points on MRI and CT surfaces are estimated, and the FEM model is applied again to deform the mesh of the MRM breast image for the second time. This step is followed by creation of a second warped MRM breast image. This time, the MRM and PET images match very well (Fig. 5, right panel). Finally, nonrigidly registered MRM and PET images are fused (Fig. 6).
 


Fig. 6. Registered and fused PET and MRM breast images. Yellow: PET. Gray: MRM. The suspicious lesion area is circled.
                (click on image to see larger view)
 

Related Publications
Ioana Coman, Andrzej Krol, David Feiglin, Wei Lee, Edward Lipson, James Mandel, Karl Baum, Mehmet Unlu, “Intermodality Nonrigid Breast-Image Registration”, Proceedings of  IEEE International Symposium on Biomedical Imaging, ISBI 2004 From Nano to Macro, pp.1439-1442 (2004) - link to file

A. Krol, M. Z. Unlu, K. G. Baum, J. A. Mandel, W. Lee, I. L. Coman, E. D. Lipson, and D. H. Feiglin. MRI/PET Nonrigid breast-image registration using skin fiducial markers. Physica Medica (2005) -  link to file