Mirabela Rusu1, Elizabeth A. Morris2, Elizabeth J. Sutton2, and Ileana Hancu1
1GE Global Research, Schenectady, NY, United States, 2Memorial Sloan Kettering Cancer Center, New York, NY, United States
Synopsis
Lesion identification in MR-guided biopsy exams can
be hampered by many factors, including large deformations and limited tissue
perfusion due to breast compression. Multiple post-contrast scans, image
subtraction and maximum intensity projection map generation may be needed to
relocate the lesion. This preliminary study suggests that non-rigid
registration between the (uncompressed breast) pre-biopsy series and the (compressed
breast) biopsy series may facilitate fast and accurate lesion (re)localization,
even with limited/absent lesion enhancement. Introduction
MRI-guided breast
biopsies can be lengthy procedures with false negative rates that are unacceptably high
1.
The time to (re-)locate the lesion can unnecessarily lengthen the procedure. Various
factors may affect lesion identification
2. For example, the pre-biopsy and biopsy MRIs are often acquired in different orientations, with different fields
of view and spatial resolutions. More importantly, breast compression (needed in the biopsy exams) may distort the anatomy and can limit perfusion, leading
to reduced or inexistent lesion enhancement in the biopsy series. Numerous post-contrast
scans, image reformatting, image subtraction and maximum intensity projection
map generation may be needed to localize lesions, thus lengthening the procedure
and increasing patient discomfort. In this proof-of-concept study, we investigate
whether accurate alignment of the pre-biopsy and biopsy images would enable one
to project the lesion identified in the pre-biopsy exam to the biopsy series,
thus simplifying lesion identification during biopsy. We introduce a registration
methodology to align the pre-biopsy exam (showing both uncompressed breasts) with
the biopsy exam (depicting only one compressed breast), and quantify its
performance in a small in vivo study.
Methods
Our study includes pre-biopsy and biopsy MRIs from four
subjects: 1 normal control (no contrast) and 3 biopsy proven breast cancer
subjects. The 1/3 sets of images were acquired on a 3T/1.5T (MR750/MR450) GE MR
scanners, respectively. Our approach for aiding lesions re-localization is
based on the automated registration of the pre-biopsy MRI to the reference
biopsy MRI; its three steps are summarized in Figure 1 and detailed below:
-
First, the MRI
intensities for the pre-biopsy and biopsy MRIs are standardized, to remove the
irregular enhancement caused by the contrast agent or other heterogeneous
illumination artifacts
3.
-
Next, an
iterative grid-search is performed to best align the two sets of images based
on their mutual information. In this search, five parameters (the X/Z scaling
factors and three translations) are optimized using an exhaustive exploration
of the parameter space. In the first iteration, a coarse grid is chosen to identify
the spatial translation of the (dual-breast axial) pre-biopsy images relative to
the (single breast, sagittal) biopsy data set. The following iterations are
executed with finer grids and seek to refine the five parameters.
-
Finally, a
multi-resolution free form deformation
4 is performed to elastically
register the pre-biopsy and biopsy scans. The grid search and deformable
registration are performed using the ITK-based package elastix
5.
The registration accuracy was estimated by measuring
the root mean squared deviation (RMSD) of anatomic landmarks, e.g. breast edge,
blood vessels etc. For each study, 3-4 anatomic landmarks were manually
identified, for a total of 14 landmarks for the four cases. The overlay between
the actual and predicted lesion location was quantified using Dice similarity coefficients.
Results
The three-step process resulted in very good alignment
of the pre-biopsy and biopsy scans in all subjects. Figures 2a-b shows an
exemplary initial alignment after the iterative multi-grid registration (step
2); Figures 2c-d demonstrates the final, refined result, after deformable
registration (step 3). Note the exquisite alignment between the two image sets,
in which one’s features extend seamlessly into the second (see arrows). Figure
3 shows an example of the overlap between the actual lesion from the biopsy MRI
(red) and of the predicted lesion location, obtained through the registration
of the pre-biopsy and biopsy data sets (blue). This result suggests that a
process similar to the one employed here may render the administration of the
contrast agent in the biopsy exam unnecessary.
Our procedure works very well in large deformation
conditions; following scaling and translations, compressions/expansions on the
X/Z axis were as large as 35%/30%. The RSMD of the 14 landmarks (summarized in
Figure 4 as a function of registration step) was 12.3 mm after the iterative
grid search, further improved by the multi-resolution deformable registration to
a 3.5mm. The average Dice coefficient between the predicted/actual lesion locations
was 59%, on account on small lesion size and difficulty in accurately outlining
lesion boundaries.
Conclusions
We developed an approach for aligning the pre-biopsy and biopsy breast MR images, enabling the easy identification of breast lesions in biopsy exams. Our final landmark
RMSD (3.5mm) approaches the slice spacing in our biopsy exams (3mm). It is
likely that higher resolution biopsy images will further increase the
achievable precision. Our predicted
lesion location (situated within a few pixels of the actual enhancing lesion)
will enable clinicians to quickly and accurately locate the lesions in the
biopsy exams, thus shortening procedure time, reducing cost, and minimizing patient
discomfort.
Acknowledgements
This work was supported in part by NIH grant
1R01CA154433References
[1] Bahrs et al, Clin Radiol
2014;69(7):695-702.
[2] Kinkel K et al, J. Mag Reson Imag 2001;
13:821-829.
[3] Cohen et. al. “Rapid and effective correction
of RF inhomogeneity for high field magnetic resonance imaging,” Hum. Brain
Mapp. 10, 204–211 (2000).
[4] Rueckert, Daniel, et al. "Nonrigid
registration using free-form deformations: application to breast MR
images." Medical Imaging, IEEE Transactions on 18.8 (1999): 712-721.
[5] Klein,
Stefan, et al. "Elastix: a toolbox for intensity-based medical image
registration." Medical Imaging, IEEE Transactions on 29.1 (2010): 196-205.