quantitative data between individuals in a highly auto ... Two schools of thought currently exist with re gard to PET image .... J Cereb Blood Flow Me/ab, Vol.
Journal of Cerebral Blood Flow and Metabolism 11:A51-A56 © 1991 The International Society of Cerebral Blood Flow and Metabolism
Region of Interest Issues: The Relationship Between Structure and Function in the Brain *tJohn C. Mazziotta, tCharles C. Pelizzari, tGeorge T. Chen, §Fred L. Bookstein, and t Daniel Valentino *Department of Neurology, Division of Nuclear Medicine and tDepartment of Radiologic Sciences, UCLA School of Medicine and Laboratory of Nuclear Medicine, Los Angeles, California, :f:Department of Radiation Oncology, Michael Reese/University of Chicago Center for Radiation Therapy, University of Chicago, Chicago, Illinois, and §Center for Human Growth and Development, University of Michigan, Ann Arbor, Michigan, U.S.A.
Summary: The comparison of data sets from individual subjects between imaging modalities is necessary in or der to evaluate the normal physiologic responses of the brain or the pathophysiological changes that accompany disease states. Similarly, it is critical to compare data between individuals both within and across imaging mo dalities. In a collaborative project with a number of uni versity groups, we have developed a system that allows for the within-subject alignment and registration of three dimensional data sets obtained from different modalities for the same individuals. These data make use of pro posed criteria for the optimal solution to positron emis sion tomography image acquisition and analysis originally established through a series of international workshops.
The analysis takes into account errors induced by image acquisition, registration, and alignment with regard to scaling, translation, and rotation. Using the principles of morphometrics and homologous landmarks, the between subject warping of individual brain anatomy to match that of other individuals, groups or an idealized model can be obtained. Resultant information can provide averaged be tween-subject data for populations of normal individuals or patients with specific neurologic disorders. Such a sys tem, provides the means by which to compare objectively quantitative data between individuals in a highly auto mated fashion. Key Words: Positron emission tomogra phy-Image correlation-Magnetic resonance imaging Morphometrics-Brain atlas.
Most approaches to the quantitative analysis of
ysis from emission tomography was completed
data from positron emission tomography (PET) re
(Mazziotta and Koslow, 1987). Results of those ef
quire its anatomical regionalization. As the disci
forts led us to develop a series of optimal criteria for
pline and field of PET has grown, so too has the
the solution of this complex problem (Table 1). In
spatial resolution of the imaging instruments, the
addition, it was recommended that investigators
number of centers and investigators, the number of
with expertise in the disciplines of physics, bio
biochemical and physiological processes that can be
mathematics, neuroanatomy, computer science,
examined, and the scope of the applications to dis
and nuclear medicine quality control join forces,
ease states. As long as 5 years ago, the complicated
share data, and exchange standardized protocols
issue of regionalization of PET data was addressed
with the goal of optimizing PET data analysis. Some
(Mazziotta, 1984), a survey of all centers perform
of these objectives have been realized.
ing emission tomographic studies was conducted,
Since 1987, a number of regionalization ap
and a series of international workshops to assess
proaches have been either proposed or imple
the goals and obstacles of data acquisition and anal-
mented (Table 2) (Bajcsy et aI., 1983; Bohm et aI.,
1983; Fox et ai., 1985; Evans et ai., 1988; Pelizzari et al., 1989). Each approach listed in Table 2 has its O'(Nn ao.vantages ano. Yunita\'\.ons. S\.nce \9'&/,
1'I.ddn:ss cones-pondence and re-PTlnt requests to Dr. 1. c. Mazziotta at Department of Neurology, UCLA School of Med icine, Laboratory of Nuclear Medicine, Los Angeles, CA 90024, U.S.A. PET, Abbreviations used: MRJ, magnetic resonance imaging: positron emission tomography; ROJ, region of interest.
\.\. has
become clear that hybrid combinations of these ap proaches might lead to the most optimal fulfillment of the criteria listed in Table 1. New realizations have included that a rigid portion of the human
AS]
A52
J. C. MAZZIOTTA ET AL.
TABLE 1. Proposed criteria for the optimal solution to the problem of PET image acquisition and analysis 1. Reproducible 2. Accurate 3. Independent of tracer employed 4. Independent of instrument spatial resolution 5. When possible, independent of ancillary imaging techniques 6. Minimizes subjectivity and investigator bias 7. Fixed assumptions about normal anatomy not required 8. Acceptable to subjects' level of tolerance (headholders, etc.) 9. Performs well in serial studies of the same patient and individual study of separate patients in a population 10. Capable of evolving toward greater accuracy as information and instruments improve 1 1. Reasonable in cost 12. Equally applicable in both clinical and research settings 13. Time efficient for both data acquisition and analysis Taken from Mazziotta and Koslow (\987).
body, e.g., the skull, may serve as its own fiduciary system allowing for cross-correlation of identical structures in the same subject between imaging mo dalities (Figs. 1-4) (Pelizzari et aI., 1989). Second, three-dimensional volumes of interest are more de sirable for full sampling of cerebral neuroanatomy than two-dimensional regions of interest (ROIs) (Evans et aI., 1988). Third, the combination of ste reotactic approaches (Talairach et aI., 1967) with individualized neuroanatomy, obtained from struc ture-function matching approaches, may lead to the most automated and yet individualized method available. Two schools of thought currently exist with re gard to PET image analysis. Some groups have con centrated their efforts on analysis of patterns or re sponses between studies in the functional PET im ages alone without regard to neuroanatomy. These productive efforts have resulted in improved schemes to identify small, both in magnitude of re sponse and in spatial distribution, signals or differ ences between studies in PET data. Other groups have produced schemes to merge structural mag netic resonance imaging (MRI) and PET data. Ulti mately, these two approaches that complement TABLE 2. Proposed or implemented ROJ analysis schemes 1. Manual template matching 2. Semiautomatic "elastic" matching of templates with x-ray CT and MRI 3. Automatic structure (x-ray CT or MRI)-function (PET) matching 4. Stereotaxic methods based on Talairach et al. ( 1967) 5. Nonanatomical statistical approaches 6. Hybrid combinations of the above For methods 1 and 4, geometric (e.g., oval, square, and circle) or irregular ROIs may be used.
J Cereb Blood Flow Metab, Vol, 11, Supp/' 1, 1991
each other will be linked in systems that detect sub tle patterns or responses in three-dimensional PET data and can also identify, on companion or ideal ized structural systems, the precise neuroanatomi cal locations of these patterns. A
CURRENT SYSTEM UNDER DEVELOPMENT
A system that performs both within-subject reg istration of image data sets and between-subject correlations is currently being developed through a collaboration among UCLA, the University of Chi cago, and the University of Michigan. This ap proach employs the logic sequence depicted in Fig. 1 and uses two separate processes, each termed a merger. Merger 1 Merger 1 results in the alignment and registration of data from different imaging modalities within the same subject using modifications of the software developed by Pelizzari et ai. ( 1989). Typically, MRI and PET data are used. These data should include both transmission and emission PET images and the MRI should be highly sampled along the Z axis us ing a pulse sequence that maximizes gray-white and brain-CSF boundaries (e.g., TE 30 ms, TI 300 ms, TR 1,276 ms). Both PET and MRI data must be corrected for distortions induced by the image acquisition pro cess itself. To this end, phantom studies have been performed identifying a variety of distortions in both the MRI and the PET data sets. After correc tion for distortions, a surface that can be visualized in both scans is segmented by contouring on each slice of both scans. The surface used may be the scalp, which is visualized on PET transmission im ages, or the brain surface, which is seen in many emission studies. For blood flow, metabolism, or other studies in which the emission data are distrib uted throughout the brain, use of the brain surface is possible. This allows use of image data from sys tems that do not produce transmission images. In addition, any potential error due to misregistration of transmission and emission images is eliminated by use of a surface defined directly from emission images. Use of transmission images to define the external surface is essential for images in which the emission image has limited anatomical detail (e.g., some receptor studies). The contours extracted from the image slices are stacked together using their known slice positions to form two three-dimensional models of the sur face as visualized in the respective scans. The University of Chicago software is used to find the coordinate transformation (including three=
=
=
A53
REGION OF INTEREST ISSUES MERGER
MERGER 2
1
BETWEEN SUBJECTS
WITHIN SUBJECT
FIG. 1. Two step process for within (merger 1) and between (merger 2) sub ject comparisons. Details of this process are described in the text.
Correction for Instrument Distortion
dimensional translation and rotation, scale differ ences due to image pixel size, and scanner gantry inclination if necessary) that optimally matches the two models. Phantom measurements (Pelizzari et aI., 1989) have demonstrated that the residual misfit between models of the external surface from MRI and PET transmission studies is on the order of 1.5 to 2 mm, averaged over the surface. The coordinate transformation so determined, which optimally matches the two surface models, describes the dif ference in orientation between the two scans rela tive to the subject, and can therefore be used to align and register the scans. Note that this transfor mation assumes that the surface is undeformed be tween the two scans; relaxation of this assumption is implicit in merger 2, as discussed below. Merger 1 also assumes that the coordinate transformation that optimally matches the surface as visualized in the two scans may be used to transform internal coordinates, i.e., the internal anatomy is rigidly
t
Correction for Anatomical Variability
Correction for Trans/ation and Rotation
fixed with respect to the surface. Both assumptions are well satisfied for brain imaging studies. Information transfer between scans may be ac complished by transforming points or volumes of interest from one scan to the other (e.g., anatomical structures delineated on MR images may be trans formed to the space of the PET images and used as a subject-specific template for PET data analysis), or new images may be produced by resampling the image data of one scan along the planes of the other (e.g., new MRI slices precisely registered with the planes of a PET study may be produced). Phantom measurements indicate that an accuracy on the or der of I to 2 pixel sizes (of the scan with the more limited resolution) may be achieved in transforming coordinates of interest from one scan to another (Pelizzari et aI., 1989). Thus, for PET images with pixel size of 2.5 mm, transformation accuracy of 4 mm or better may be expected (Pelizzari et aI., 1989).
Original Data FIG. 2. Raw data from MRI and PET used as input for merger 1. MRI data were acquired at 4.2 mm intervals with a slice thickness of 4.0 mm using an inversion recovery pulse sequence (TE 30 ms, TI 300 ms, TR 1,276 ms). Both transmission and emission PET data were also acquired, although only the emission data are shown in this image set. Note the misalignment of sections in terms of scaling, angle, and level. Compare these data with the reformatted image sets seen in Fig. 4 following merger 1 process ing. =
3.
FU '
ISS!
=
=
J Cereb Blood Flow Me/ab, Vol. 11, Suppl. 1, 1991
A54
1. C. MAZZIOTTA ET AL.
FIG. 3. Three-dimensional surface con tour fitting with alignment and registra tion matrix. Solid lines demonstrate sur face contours of the head obtained from MRI while the dotted lines indicate the same boundary obtained from the PET transmission images. These two three dimensional data sets were fit using an iterative approach described by Pelizzari et al. (1989). The transformation matrix in dicates the disparities between the origi nal two data sets in terms of scaling, translation, and rotation for each of the three axes. Following the determination of this transformation matrix, PET data are reformatted and sliced to match the MRI data set (see Fig. 4).
One possible application of merger 1 is to pro duce an aligned and registered structure-function data set that allows ROIs or volumes of interest to be defined on the structural image and transferred automatically to the functional image (Fig. 4).
brain shapes must be distorted to fit either a com mon standard or an idealized model.
Merger 2 Merger 2 involves providing one single standard
If this process is to be performed automatically,
ized coordinate system for all the neurocranial
and have the capacity to be successful both be
forms of a data set or study, The coordinate system
tween as well as within a given subject, individual
FIG. 4. Aligned and registered MRI and PET emission data set following registration and alignment as described in Fig. 3. These are the same data initially depicted in Fig. 2. Note the close concordance of the structural and functional anatomy with particular refer ence to the rectal gyri of the frontal lobes, the inferior temporal lobes, the transition from cerebellum to occipital lobe, and se lected gyri.
J Cereb Blood Flow Metab, Vol. 11, Suppl. I, 1991
that is used may be thought of as the shape of a
A55
REGION OF INTEREST ISSUES standard or "atlas." This coordinate system is then deformed or warped to fit the forms observed case by case. The description of all forms as warpings of a fixed atlas serves three separate purposes: (a) quantification of the forms for all the cases of a study: this permits the rigorous biostatistical explo ration of group differences, time trends, dose response relations, or correlations with clinically measured functions; one application is to generate the atlas itself as a sample average rather than a haphazardly acquired single specimen; (b) stan dardization of all of the grey-scale images of a func tional series to a fixed geometry: this is accom plished by executing the appropriate "unwarping" on each picture; it then becomes possible to realize the grey-scale information as a well-defined feature space of its own, on which one may average pic
FIG. 5. Three-dimensional surface rendering of both struc ture and function. Outer boundaries of the skin were ob tained from the MRI data set and reconstructed in three di mensions. The cortical surface of the brain was extracted from the outer boundary of an FOG/PET emission study and thus represents cerebral glucose metabolism in color scale. These data were assembled and displayed on a PIXAR image processing system to allow for the identification of cortical surface metabolic anatomy.
tures and correlate their features with various pre dictors, clinical measures, or sequelae; and (c) cor relations of anatomical and metabolic images ("form function"): standard statistical methods can
been proposed and employed by Fox and col leagues ( 1985).
be applied to correlate geometry, as expressed in the warping function of purpose (a), with subspaces of grey-scale f�atures after warping to constant form as in purpose (b). These warping functions, called "homology maps" in the biomathematical literature, go beyond the rigid motions of merger 1. Generally, they take the form of exact interpolants between scattered geometric features that are "known" to correspond between forms. One very useful algebraic form for this interpolant is the thin-plate spline developed in the literature of surface interpolation and modified to apply to warping by Bookstein ( 1989). Extensions of this technique to more complex data, such as curves on surfaces, are a primary thrust of our morphometric research at this time. In general, a warping cannot preserve both densities and counts; the issue of what to hold invariant un der the transformation will await considerable ex perience with correlates of images after deforma tion. Three-dimensional surface renderings (Fig. 5)
of the human cortex will be used to evaluate these results and for the identification of surface land marks on the cerebral cortex. Simultaneous with the acquisition of data for merger 1 and its transformation with merger 2, ste
CONCLUSION As has been previously stated (Mazziotta, 1984; Mazziotta and Koslow, 1987), functional and struc tural anatomy are fundamentally different entities. PET investigators have always related functional images to discrete anatomical brain regions. As spa tial resolution continues to improve, this is done with increasing confidence. Yet, in reality, these images represent functional data that are not equiv alent to structural information. Many tracers, par ticularly those that bind to specific neurochemical subsystems of the brain (e.g., eSF]fluorodopa,
e1C]methylspiperone), produce images with low structural information content. Thus, the idea that increasing spatial resolution in PET will solve the image analysis problem is erroneous. Complex
analysis systems, such as those described in this report, will undoubtedly be employed to regionalize and standardize PET data analysis in the most op timal format. Realization and acceptance of these objectives will require a continued concerted effort among PET investigators so as to maximize the power of these functional imaging techniques to ex plore brain function in health and disease.
reotactic (Talairach et ai., 1967; Bohm et ai., 1983;
Fox et ai., 1985) and statistical (Clark et ai., 1985)
approaches can also be applied to these data sets. This opportunity should provide a good test of the validity of a wide variety of regionalization meth ods, a goal that is intended as part of this project. Ultimate validation will also include prediction and localization of functional stimuli, as has already
Acknowledgment: This work was supported, in part, by NIMH Grant #MH37916, NINDS Grant #NS15654, and DOE Contract #DEAC03SF7600012. The authors wish to acknowledge the efforts of H. K. Huang, D.SC., Duffy Cutler, Ed Hoffman, Ph.D., and Scott Grafton, M.D., and to acknowledge gratefully the efforts of Lee Griswold in the preparation of the figures and Maureen Chang and Maggie Marquez for the preparation of this manuscript.
J Cereb Blood Flow Metab, Vol. 11, Suppl. 1, 1991
A56
J. C. MAZZIOTTA ET AL. REFERENCES
Bajcsy R, Lieberson R, Reivich M, Berggren BM, Olsson L (1983) A computerized system for the elastic matching of deformed radiographic images to idealized atlas images. J Comput Assist Tomogr 7:618-625
Bohm C, Greitz T, Kingsley D, Berggren BM, Olsson L (1983) Adjustable computerized stereotaxic brain atlas for trans mission and emission tomography. Am J NeuroradioI4:731733
Bookstein FL (1989) Principle warps: Thin plate splines and the decomposition of deformation. IEEE Trans Pattern Anal Med Intelll1:567-585
Clark C, Carson R, Kessler R, Margolin R. Buchsbaum M, DeLisi L, King C, Cohen R (1985) Alternate statistical mod els for the examination of clinical data positron emission tomography/fluorodeoxyglucose data. J Cereb Blood Flow Metab 5:142-150
Evans AC, Beil C, Marrett S, Thompson Cl, Hakim A (1988)
J Cereb Blood Flow Me/ab. Vol. 11. Suppl. 1, 1991
Anatomical-functional correlation using an adjustable MRI based region of interest atlas with positron emission tomog raphy. J Cereb Blood Flow Metab 8:513-530 Fox PT, Perlmutter lS, Raichle ME (1985) A stereotactic method of anatomical localization for positron emission tomogra phy. J Comp Assist Tomogr 9:141-153 Mazziotta lC (1984) Physiologic neuroanatomy: New brain im aging methods present a challenge to an old discipline. J Cereb Blood Flow Metab 4:481-483
Mazziotta lC, Koslow SH (1987) Assessment of goals and ob stacles in data acquisition and analysis from emission tomog raphy: Report of a series of international workshops. J Cereb Blood Flow Metab 7(suppl 1):SI-S31 Pelizzari CA, Chen GTY, Spelbring DR, Weichselbaum RR, Chen CT (1989) Accurate three-dimensional registration of CT, PET and/or MR images of the brain. J Comput Assist Tomogr 13:20-26
Talairach 1, Szikla G, Tournoux P (1967) Atlas d'anatomie ste reotaxique du telencephale, Paris, Masson
