utililization of experimental animal model for correlative multispectral mri and pathological...

PII S0730-725X(99)00090-9

● Original Contribution

UTILILIZATION OF EXPERIMENTAL ANIMAL MODEL FOR CORRELATIVEMULTISPECTRAL MRI AND PATHOLOGICAL ANALYSIS OF BRAIN TUMORS

JENNIFER GORDON,* FEROZE MOHAMED,† SIMON VINITSKI,‡ ROBERT L. KNOBLER,§ MARK CURTIS,¶SCOTT FARO,† AND KAMEL KHALILI *

*Center for Neuro Virology and NeuroOncology and †Department of Radiology, MCP Hahnemann University, Philadelphia, PA19102; and Departments of ‡Radiology, §Neurology, and ¶Pathology, Anatomy, and Cell Biology, Jefferson Medical College,

Thomas Jefferson University, Philadelphia, PA 19107, USA

Magnetic resonance imaging is the method of choice for non-invasive detection and evaluation of tumors of thecentral nervous system. However, discrimination of tumor boundaries from normal tissue, and the evaluation ofheterogeneous lesions have proven to be limitations in traditional magnetic resonance imaging. The use ofpost-image acquisition processing techniques, such as multispectral tissue segmentation analysis, may providemore accurate clinical information. In this report, we have employed an experimental animal model for braintumors induced by glial cells transformed by the human neurotropic JC virus to examine the utility ofmultispectral tissue segmentation for tumor cell identification. Six individual tissue types were discriminated bysegmentation analysis, including heterogeneous tumor tissue, a clear demarcation of the boundary betweentumor and non-tumor tissue, deep and cortical gray matter, and cerebrospinal fluid. Furthermore, the segmen-tation analysis was confirmed by histopathological evaluation. The use of multispectral tissue segmentationanalysis may optimize the non-invasive determination and volumetric analysis of CNS neoplasms, thus providingimproved clinical evaluation of tumor growth and evaluation of the effectiveness of therapeutic treatments.© 1999 Elsevier Science Inc.

Keywords:Tissue segmentation; Histopathology; JCV; T-antigen.

INTRODUCTION

Conventional magnetic resonance imaging (MRI) isbased on the detection of hydrogen atoms (protons),which can provide excellent contrast between differenttypes of tissues in the human body.1 In addition, the dataobtained during imaging can be further processed usinga variety of methods including filters and statistical anal-ysis. In this manner, post-processing techniques can ex-tract further information from the acquired image thanthat obtained by conventional MRI alone and providemore information on the composition, boundaries ormargins, and volume of the affected region.

High grade brain tumors of glial origin have a poorprognosis for long term survival and are especially dif-ficult to manage, as they are extremely aggressive andoften require surgical resection and subsequent radiationtherapy.2 Current techniques for brain tumor diagnosis

and treatment rely heavily on MRI. This technique, how-ever, has not been well correlated with pathologic data inevery case, as evidenced by poor discrimination betweenseveral pathologic processes including infarction, neo-plasms, inflammation, and the state of myelination. Fur-thermore, MRI may not accurately discriminate acutelesions from chronic, or precisely define the bordersbetween neoplastic and surrounding normal tissue.

The technique of magnetic resonance multispectraltissue segmentation analysis can be an extremely usefulmethod for investigating brain tumors in vivo followingchemotherapy and radiation therapy. A number of seg-mentation techniques have been developed in order toidentify brain pathology in the clinical setting,3 includingtumors, but histologic identification of tumor tissueswithin and surrounding the tumor has not been possiblein patient studies. Since MRI is the imaging procedure of

RECEIVED 1/16/99; ACCEPTED8/5/99.Address correspondence to Dr. K. Khalili, Center for Neu-

rovirology and Cancer Biology, Temple University, 1900North 12th Street, Room 203, Philadelphia, PA 19122.

Magnetic Resonance Imaging, Vol. 17, No. 10, pp. 1495–1502, 1999© 1999 Elsevier Science Inc. All rights reserved.

Printed in the USA.0730-725X/99 $–see front matter

1495

choice to study human brain tumors, it is critical tocompare segmentation data with the composition of liv-ing biologic tissue to validate the post processing tech-niques.

We have developed a defined volumetric three dimen-sional (3D) multispectral tissue segmentation technique,which is able to precisely identify and characterize thegrowth of the brain tumor and surrounding tissue. Here,we have utilized an animal model of glioblastoma in-duced by intracerebral inoculation of newborn GoldenSyrian hamsters with cells transformed by the humanpolyomavirus, JCV. JCV is a neurotropic virus which isthe causative agent of the fatal demyelinating disease,Progressive Multifocal Leukoencephalopathy (PML), inimmunocompromised individuals.4 This virus, when in-oculated into hamster neonates, induces glial origin tu-mors, such as glioblastoma multiforme and astrocyto-mas, as well as medulloblastomas.5,6 Moreover, glialcells derived from the JCV-induced tumor which expressthe viral oncoprotein, T-antigen, have the ability to in-duce tumors 3–6 months after intracerebral inoculationof newborn hamsters. Using this model, we have corre-lated information from 3D multispectral tissue segmen-tation with post-mortem histopathological analysis ofbrain tumors induced in experimental animals. This in-formation provides a topographical map with specifictumor distribution.

MATERIALS AND METHODS

Induction of Tumors in Neonatal HamstersNewborn Golden Syrian hamsters were intracere-

brally inoculated approximately 48 h after birth in theright cerebral hemisphere with 105 cells of the JC virus-transformed HJC-15c cell line in a volume of 10mL viaa Hamilton syringe with a 26 gauge needle.7 The animalsexhibited progressively deteriorating motor coordinationand righting reflexes but were otherwise healthy.

Data Acquisition and Image Post ProcessingAnimals exhibiting neurologic abnormalities were se-

lected for imaging, which was performed under generalanesthesia, 1–2% isofluorane, administered via a nosecone. The small brain of the hamster (approximately 1.5cm in diameter) produced an inherently low signal-to-noise ratio (SNR) which provided challenges to thestudy. In order to obtain proper data for image post-processing, three criteria were maximized: high resolu-tion of the images, high SNR, and high tissue contrast.Animals were imaged in a GE 1.5 Tesla conventionalclinical MRI scanner utilizing a conventional wrist coil.A detailed account of the MR imaging techniques hasbeen described elsewhere.1 In brief, all MR imaging wasperformed using a 3D volume acquisition mode to max-

imize SNR. High resolution MR images were obtainedwith a voxel size of 0.1 mm.3 Three types of MR imagingwere performed; gradient echo (GRE), inversion recov-ery gradient echo (IR-GRE), and steady state free pre-cession (SSFP), each of which produced different tissuecontrast. Imaging parameters were: A) GRE: TR/TE520/6.8 ms, 8 averages (number of acquisitions); B)SSFP: TR/TE5 48/75 ms, 8 averages; and C) IR-GRE:TR/TE 5 20/6.8 ms, TI5 500 ms, 6 averages. Multi-spectral segmentation was done with the three MR im-ages described above used as inputs. Following MR dataacquisition, images were transferred to a Sun Sparc Sta-tion for multispectral tissue segmentation analysis. Afterthe calculation of a 3D feature map, a stack of color-coded images was created, arbitrarily assigning colors tothe various segmented areas. The multispectral tissuesegmentation technique is described elsewhere.8 Briefly,RF inhomogeneity is minimized by a technique sug-gested by Mohamed et al.9 and extended for multispec-tral tissue segmentation. Nonlinear “diffusion type” fil-ters10,11 are applied to reduce random fluctuations ofnoise from the images. Next, the qualified observer (neu-roscientist) “seeds” tissue samples (40–50 samples/tis-sue), and the k-Nearest Neighborhood (k-NN) segmen-tation is utilized for both 2D and 3D feature mapcalculations. This choice allowed searching for the mostdense cluster of each tissue sample in 3D feature spaceand utilized only those samples as input for segmenta-tion.12 A connectivity algorithm13 along with a dividingcube algorithm14 was used to construct a surface ofselected tissue(s).

Histopathology and ImmunohistochemistryAnimals were sacrificed immediately following MRI.

For histology, the intact skulls were removed, decalcifiedwith Decal II (Surgipath), and coronal sections weremade. Sections were stained with hematoxylin and eosin(H & E) for light microscopy according to standardprotocols.15 For immunohistochemistry, animals wereperfusion fixed, the brain and tumor were removed andparaffin embedded, and immunostaining of adjacent par-affin sections with mouse monoclonal antibody to SV40T-antigen which is cross reactive with JC virus T-antigen(Oncogene Science Ab 416) was performed by the Vec-tor Elite ABC system and detected with diaminobenzi-dine tetrahydrochloride (DAB) as described previous-ly.16

RESULTS AND DISCUSSION

The animal model is produced by intracerebral injec-tion of the JCV-induced glioma cell line, HJC-15c, intothe right cerebral hemisphere of newborn Golden Syrianhamsters.7 Three to six months after injection, hamsters

1496 Magnetic Resonance Imaging● Volume 17, Number 10, 1999

harboring tumors were subjected to MR imaging whileunder anesthesia. Following imaging, the hamsters weresacrificed and the brain tissue was removed for neuro-pathological and histologic analyses. Animals inoculatedwith JC virus-transformed cells had clearly identifiabletumor masses as compared with normal age-matchedcontrol animals (Fig. 1A and B). Coronal sections of theheads of both animals are shown at the level of thethalamus. The coronal section of treated hamster brain,skull, and surrounding soft tissue demonstrates a tumorlocated on the superior lateral surface of the cerebralhemisphere. The tumor mass compresses the right andleft hemispheres toward the base of the skull with devi-ation of the central fissure laterally. The border betweenthe tumor mass and the brain appears distinct and noinfiltration of brain parenchyma by the tumor was iden-tified grossly or microscopically. The tumor also erodedthrough the superior aspect of the skull and extended into

extracranial soft tissue. Extensive tumor necrosis wasidentified in the central portion of the tumor.

Histopathological analysis revealed the tumor to behypercellular and composed of cells with fibrillary ap-pearing cytoplasm and medium sized to large, ovoidnuclei with irregular contours (Fig. 1C). Giant tumorcells were present. Immunohistochemical stains for in-termediate filaments glial fibrillary acidic protein(GFAP) and vimentin demonstrated positive cytoplasmicstaining in tumor cells (data not shown). Numerous mi-totic figures were present throughout the neoplasm andlarge areas of tumor necrosis were present. The histologyis consistent with a poorly differentiated malignant tu-mor of glial origin. Immunohistochemical stains for theneuronal markers synaptophysin and neurofilament werenegative in the tumor (data not shown). Tumor tissue wasevaluated for the presence of the JC virus transformingprotein, T-antigen. As shown in Fig. 1D, immunostain-

Fig. 1. Induction of CNS tumors in newborn Golden Syrian hamsters by JC virus T-antigen transformed cells. Gross coronal sectionthrough the brain, tumor, and skull of treated hamster H-046 at six weeks of age (A); gross coronal section of normal age-matchedcontrol, hamster H-053 (B); hematoxylin and eosin staining of paraffin-embedded tumor tissue from treated littermate hamster H-047(C); and immunohistochemical staining of adjacent section from treated hamster H-047 with antibody to JC virus T-antigen detectedwith DAB and counterstained with hematoxylin demonstrates the presence of JC virus T-antigen (original magnification3400) inthe nucleus of tumor cells (D).

Correlative MRI and pathological analysis● J. GORDON ET AL. 1497

ing of the tumor revealed substantial nuclear staining forJCV T-antigen.

Due to the small size of the hamster brain, volume 3DMRI sequences were used to increase image resolution.The inherently low signal-to-noise ratio provided chal-lenges to the study. In order to obtain proper data forpost-processing of images, three parameters were maxi-mized: high resolution of the images, high signal-to-noise ratio, and tissue contrast. The animals were imagedin a GE 1.5T MRI unit utilizing a wrist coil, the smallestconventional transmitter-receiver available, to maintainRF uniformity. The images shown in Fig. 2A–C repre-sent those obtained with the pulse sequences GRE, IR-GRE, and SSFP, respectively. Following acquisition, theimages were transferred to a Sun Sparc Station for mul-tispectral tissue segmentation and post-processing was

carried out as described in the Materials and Methods.The process of seeding involves the identification ofdifferent components of tissue samples with the aid of acrosshair marker that allows for the simultaneous selec-tion of a voxel of interest from all three of the MR imagesequences. This seeding results in grouping of clusters oftissue types in three dimensions, as demonstrated in Fig.2D. These data are then used as the input for k-NNresulting in the separation of points into distinct tissuesas schematized in Panel E. After segmentation, a stack ofcolor-coded segmented images is created where up to sixtissues were classified, which can be visualized in twodimensions (Fig. 2F).

In order to further examine the nature of the tissuesidentified by MRI and segmentation analysis, animalswere sacrificed immediately following imaging and the

Fig. 2. Magnetic resonance image data acquisition, and image post processing/segmentation analysis of hamster brain tumor.Hamster H-036 was imaged in a GE 1.5 Tesla MR imager using a small transmitter-receiver surface coil. Three types of MR imagingwere performed, as described in the Materials and Methods. The presented images represent the same anatomic location: gradientecho imaging (GRE) with RF spoiling (A); steady-state free precession (SSFP) (B); inversion recovery gradient echo (IR-GRE)images. Image post-processing followed by tissue classification using a multispectral tissue segmentation method, an algorithm whichseparates and groups (clusters) tissues with similar MR signal characteristics, was performed (C); plotting of tissue points based onthe three input MR images GRE, SSFP, and IR-GRE demonstrating clusters of segmented points (D); schematic illustrating separationof points into distinct tissue types based on input data as shown in panel D. S1, S2, S3 represent three different MR signal intensities(E); and segmented data are used to create a stack of color-coded images that separates each type of tissue by color and shows preciselocation (F). The following tissues or structures were identified: black, CSF; red, pink, and brown, heterogeneous tumor; green,cortical gray matter; yellow, deep gray matter.


skulls of the animals were subjected to histopathologicalevaluation. As shown in Fig. 3, input data from the 3Dvolume acquisition were utilized to separate tissue clus-ters by multispectral tissue segmentation (Fig. 3A and B,respectively). A significant degree of resolution can beobserved with the post process analysis shown in Figs. 2and 3, bearing in mind the small size of the hamster brain(approximately 1.5 cm in diameter). Histopathologicalanalysis of the same hamster brain, skull, and soft tissuesof the head revealed a tumor mass on the superior aspectof the right cerebral hemisphere, as shown in Fig. 3C.Multispectral tissue segmentation analysis allowed us toeasily distinguish between several normal tissues withinthe brain including deep gray matter, cortical gray mat-ter, and cerebrospinal fluid (CSF). Additionally, the tu-mor tissue was identified as containing more than onecomponent, most likely due to the presence of necrotictissue as well as edematous or hemorragic tumor. Excel-lent correlation was obtained between the segmentationand the macroscopic and microscopic appearances of thetumors. The information obtained from both segmenta-tion and histologic analysis visually presented in Fig. 3D.

Of note, within the portion of the cortex visualized inFigs. 2 and 3, the majority of the hamster brain iscomposed of cortical gray matter and the deep nuclei ofgray matter including the corpus striatum and the thala-mus. Solid white matter in these sections would beconfined to the corpus callosum, as shown in Fig. 2D,which may not be present in significant amounts to bedetected with the imaging and segmentation parametersutilized in this study. Further, the presence of occasionalbrown, pink or red pixels, which are assigned to hetero-geneous tumor tissue, can be seen outside the boundaryof the tumor and appear to be localized within the normalbrain parenchyma. On closer inspection, one may ob-serve that these pixels in fact localize to the lateral orthird ventricles, perhaps indicating the presence of ahemorrhage or edema that has accumulated within theventricular system.

In comparing tissue segmentation data and histopa-thology, it is evident that the multispectral approachallows for a more clear distinction between tumor andnon-tumor tissue. Further, a distinction between differentnormal tissues such as deep and cortical gray matter can

Fig. 3. Correlation of MR images, multispectral tissue segmentation analysis, and histopathology. MR image of hamster H-046obtained by SSFP method (A); segmented image of hamster H-046 obtained from three types of MR images. The following tissuesor structures were identified: black, CSF; red, pink, and brown, heterogeneous tumor; green, cortical gray matter; yellow, deep graymatter (B); hematoxylin and eosin histologic staining of coronal section of hamster brain, and soft tissue of the head (C), after paraffinembedding and sectioning of sample from hamster H-046 shown in Fig. 1A. The tumor mass is located in the superior aspect of theright cerebral hemisphere; and Camera lucida demonstrating anatomic structures identified by both imaging and histology (D).


be made more readily than by traditional MRI. Of par-ticular interest, the line of demarcation between the tu-mor and normal brain is well-characterized by the mul-tispectral tissue segmentation analysis, an importantfeature as tumor boundaries are extremely difficult topredict by conventional MRI alone. Finally, the tumoritself demonstrates several clusters of tissue, whichwould correspond well with the heterogeneous nature ofthe tumor, as observed histopathologically. Additionalstudies may allow for enhanced discrimination betweenentities such as edema, necrosis, hemorragic tumor, andnormal CSF.

Conventional MR images alone are not capable ofaccurate tissue type segmentation and volumetric analy-sis. Several techniques have been under investigation foraccurate classification of the tissues obtained from MRimages.17 A wide variety of approaches have been used,ranging from totally manual to automated techniques forvolumetric analysis of MR images. Manual methods areentirely based on prior clinical knowledge of the anat-omy in the image by the observer, and are time consum-ing. Hence, they are subject to higher interobserver andintraobserver variability. However, with automatedmethods, the reproducibility of the results can be assuredand are less time consuming.

Several methods of automated classification such asmultispectral tissue classification, seed growing meth-ods, fuzzy connectedness, global thresholding, sterotac-tic transformation, and local thresholding methods arecurrently under investigation particularly for tissue clas-sification of multiple sclerosis (MS) lesions.17 Most ofthese methods utilize two input images (i.e., T1 andProton density) for multispectral tissue segmentation,and are also affected by uncorrected variation in signalsensitivity within slices and between slices, which canlead to misclassification of the tissue types. Further, noneof these techniques have been validated using animalmodels. However, in our multispectral tissue segmenta-tion method we have utilized three inputs, based on theassumption that tissue cluster separation will be in-creased in 3D feature space. The benefits of using mul-tiple weighted images for classification of brain tissuehas been demonstrated.18 This technique has been usedin this study for accurately classifying the brain tumortissues in the hamster model.

Our initial investigations using segmentation analysiswere based on the algorithm developed by Cline et al.19

which calculates the distribution of each tissue of interestin three dimensional space based on two conventionalimage inputs (e.g. proton density and T2-weighted). In arelated study we hypothesized that adding a third inputimage would make the separation of tissue types in threedimensional (3D) space greater, with a resulting im-provement in resolution of tissue segmentation. In a

recent study using this technique in patients with multi-ple sclerosis, we found that the addition of T1 in thealgorithm provided new information not obtained in the2D Proton density and T2-weighted sequences.8 There-fore, we believe that using three input images for seg-mentation of tumors may provide vital information re-garding the composition and extent of brain tumors.

MR imaging of normal mouse brain was performedand compared with histologic sections by Munasinghe etal.20 The imaging was performed at very high field of 4.7Tesla. Conventional T2-weighted spin echo imaging wasperformed, and the images were obtained with a resolu-tion of 0.02 mm.3 Using this microscopic imaging andknowing the anatomy of the normal brain, nine tissuesand structures were identified. However, no abnormalbrain tissues were investigated by either MRI or by tissueclassification (segmentation) techniques. In two studiesperformed at 1.5 Tesla for normal brain anatomy,21,22theresolution of the images that these investigators wereable to achieve was 0.3 mm.3 Consequently, less dis-crimination was achieved.

Abnormal rat brains have also been studies by inves-tigators.23,24 In a study of in vivo dynamic MRI ofMBP-induced Acute Experimental Allergic Encephalo-myelitis (EAE) in Lewis rats,24 relaxation rates weremeasured at the very high field of 4.7 Tesla. Contrastagent (Gd-DOTA) was used in this study to establish therole of the blood-brain barrier (BBB), and only slightdifferences in the tissue structures were observed. Ofnote, conventional T1- and T2-weighted imaging alongwith contrast material was utilized in this study. Thougha thresholding technique was used for following Gd-DOTA images, no histologic correlation was attempted.A contrast agent such as gadolinium was not utilized inour studies as recent data in the application of multispec-tral tissue segmentation to intracranial lesions includingmalignant brain tumors25 has shown that areas enhancedby gadolinium do not coincide with areas altered bychanges in T1 and T2 signals which are intrinsic proper-ties of the biologic tissues. In other words, enhancementdue to gadolinium occurs in selective areas where thereis a disruption in the blood brain barrier, and is notnecessarily characteristic of all the tissue within thetumor area.

In another study, histopathological correlation of MRimaging parameters in experimental cerebral ischemia23

was performed. In this study conventional T1- and T2-weighted imaging along with diffusion imaging wasused. While relaxation rates are poor predictors of animalpathogenesis, the diffusion coefficient showed some dif-ferences in the brain image studied. However, image postprocessing was not performed.

The ability of multispectral tissue segmentation anal-ysis to distinguish between tumor and non-tumor tissue,


as evidenced from the clear line of demarcation ob-served, represents the potential of this technique fordistinguishing the boundaries of CNS neoplasms moreaccurately than the currently available traditional MRI.Further, separation of tumor tissue into components in-cluding non-viable and viable tissue, and advancing mar-gin, will prove clinically important. Multispectral tissuesegmentation can accurately and efficiently calculate thevolume of the tumor and represents a non-invasive tech-nique for evaluating the effect of anti-tumor therapy onthe size and composition of the CNS neoplasm. We haveutilized an animal model of glioblastoma in order tocorrelate segmentation analysis data with histopathology.Well-established imaging parameters acquired in 3Dwith a readily available 1.5T clinical scanner were cho-sen in order to make the collection of input data forsegmentation analysis more readily adaptable to the clin-ical setting. Further validation of tissue segmentationmay lead to automated characterization and volumetricanalysis in order to provide a non-invasive tool to beused in the diagnosis and treatment of CNS neoplasms.

Acknowledgments—The authors thank past and present members of theCenter for NeuroVirology and NeuroOncology for their support, in-sightful discussion, and sharing of ideas; and C. Schriver for prepara-tion of this manuscript. This work was made possible by grant numberPO1 NS36466 awarded by NIH to KK.

REFERENCES

1. Stark, D.; Bradley, W. Magnetic Resonance Imaging, 2nded, Volume 1, Chapter 5. St. Louis, MO: Mosby YearBook, Inc.; 1992.

2. Fine, H.A. Novel biological therapies for malignant glio-mas. Antiangiogenesis immunotherapy, and gene therapy.Neurol. Clin. 13:827–846; 1995.

3. Clarke, L.P.; Velthuizen, R.P.; Phuphanich, S.; Schellen-berg, J.D.; Arrington, J.A.; Silbiger, M. MRI: Stability ofthree supervised segmentation techniques. Magn. Reson.Imaging 11:95–106; 1993.

4. Berger, J.R.; Concha, M. Progressive multifocal leukoen-cephalopathy: The evolution of a disease once consideredrare. J. Neuro. Virol. 1:5–18; 1995.

5. Walker, D.L.; Padgett, B.L; ZuRhein, G.M.; Albert, A.E.;Marsh, R.F. Human papovavirus (JC): Induction of braintumors in hamsters. Science 181:674–676; 1973.

6. ZuRhein, G.M. Studies of JC virus-induced nervous sys-tem tumors in the Syrian hamster: A review. In: J.L. Sever,D.M. Madden, (Eds). Polyomaviruses and Human Neuro-logical Disease. New York: Alan R. Liss, 1983: pp. 205–221.

7. Raj, G.; Gordon, J.; Logan, T.J.; et al. Characterization ofclonal cell lines derived from hamster glioblastomas in-duced by the neurotropic papovavirus, JCV. Int. J. Oncol.7:801–808; 1995.

8. Vinitski, S.; Seshagiri, S.; Mohamed, F.B.; Frazer, G.;Flanders, A.E.; Gonzalez, C.F.; Mitchell, D.G. Tissue

characterizaiton by MRI: Data segmentation using 3D fea-ture map. In: G. Vernazza, C. Braccine, (Eds). ImageProcessing Theory and Applications. Amsterdam: ElsevierScience, 1993: pp. 325–328.

9. Mohamed, F.B.; Vinitski, S.; Gonzalez, C.; et al. Imagenonuniformity correction in high field (1.5T) MRI. Proc.IEEE Eng. Med. Biol. 17:471–472; 1995.

10. Gerig, G.; Kubler, O.; Kikinis, R.; Joses, F.A. Nonlinearanisotropic filtering of MRI data. IEEE Trans. Med. Im-aging 11:221–232; 1992.

11. Perona, P.; Malik, J. Scale-space and edge detection usinganisotropic diffusion. IEEE Pattern Anal. Machine Intell.12:629–639; 1987.

12. Vintiski, S.; Gonzalez, C.; Burnett, C.; Seshagiri, S.; Or-tega, H.V.; Faro, S. Tissue segmentation in MRI as aninformative indicator of disease activity in the brain. In: L.DeFlariani, C. Vernazza, (Eds). Image Analysis and Pro-cessing, Berlin: Springer-Verlag, 1995: pp. 265–270.

13. Cline, H.E.; Dumoulin, C.L.; Hart, H.R. Jr.; Lorensen,W.E.; Ludke, S. 3D reconstruction of the brain from MRIusing a connectivity algorithm. Magn. Reson. Imaging5:345–349; 1987.

14. Cline, H.E.; Lorensen, W.E.; Ludke, S.; Crawford, C.R.;Teeter, B.C. Two algorithms for the three-dimensionalreconstruction of tomograms. Med. Phys. 15:320–327;1988.

15. Luna, L. Manual of histological staining methods of theArmed Forces Institute of Pathology, 3rd ed. New York:McGraw-Hill, 1968: pp. 32–47.

16. Rencic, A.; Gordon, J.; Otte, J.; Zoltick, P.; Khalili, K.;Andrews, D. Detection of JCV DNA sequence and expres-sion of the viral oncogene, T-antigen in brain of patientwith oligoastrocytoma. Proc. Natl. Acad. Sci. USA 93:7352–7357; 1996.

17. Evans, A.C.; Frank, J.A.; Antel, J.; Miller, D.H. The role ofMRI in clinical trials of multiple sclerosis: Comparison ofimage processing techniques. Ann. Neurol. 41:125–132;1997.

18. Vinitski, S.; Gonzalez, C.; Mohamed, F.; Iwanaga, T.;Knobler, R.; Khalili, K.; Mack, J. Improved intracraniallesion characterization by tissue segmentation based on a3D map. Magn. Reson. Med. 37:457–469; 1997.

19. Cline, H.; Lorensen, W.; Kikinis, R.; Jolesz, F. Threedimensional segmentation of MR images of the head usingprobability and connectivity. J. Comput. Assist. Tomogr.14:1037–1042; 1990.

20. Munasinghe, J.P.; Gresham, G.A.; Carpenter, T.A.; Hall,L.D. Magnetic resonance imaging of the normal mousebrain: Comparison with histological sections. Lab. AnimalSci. 45:674–679; 1995.

21. Ballon, D.; Graham, M.C.; Midownik, S.; Koutcher, J.A.Doubly tuned solenoidal resonators for small animal im-aging and spectroscopy at 1.5 Tesla. Magn. Reson. Imag-ing 7:155–162; 1989.

22. Rofstad, E.K.; Steinsland, E.; Kaalhus, O.; Chang, Y.B.;Houik, B.; Lyng, H. Magnetic resonance imaging of hu-man melanoma xenografts in vivo: Proton spin-lattice and


spin-spin relaxation times versus fractional tumor watercontent and fraction of necrotic tumor tissue. Int. J. Radiat.Biol. 65:387–401; 1994.

23. Helpern, J.A.; Dereski, M.O.; Knight, R.A.; Ordridge, R.J.;Chopp, M.; Qing, Z.X. Histopathological correlation ofnuclear magnetic resonance imaging parameters in exper-imental cerebral ischemia. Magn. Reson. Imaging 11:241–246; 1993.

24. Namer, I.J.; Steibel, J.; Poulet, P.; Armspach, J.P.; Mauss,Y.; Chambron, J. In vivo dynamic MRI of MBP-inducedacute experimental allergic encephalomyelitis in Lewis rat.Magn. Reson. Med. 24:325–334; 1992.

25. Vinitski, S.; Gonzalez, C.; Knobler, R.L.; Andrews, D.;Iwanaga, T.; Curtis, M. Fast tissue segmentation based ona 4D feature map in characterization of intracranial lesions.J. Magn. Reson. Imaging 9:768–776; 1999.


utililization of experimental animal model for correlative multispectral mri and pathological...

Documents