Increased differentiation of intracranial white matter lesions by multispectral 3D-tissue segmentation: preliminary results

Magnetic Resonance Imaging 19 (2001) 207–218

Increased differentiation of intracranial white matter lesions by multispectral 3D-tissue segmentation: preliminary results Feroze B. Mohamed, Ph.D.a,*, Simon Vinitski, Ph.D.b, Carlos F. Gonzalez, M.D.a, Scott H. Faro, M.D.a, Fred A. Lublin, M.D.d, Robert Knobler, M.D.c, Juan Esteban Gutierrez, M.D.e a

Department of Radiology, MCP/Hahnemann University, Philadelphia, PA, USA Department of Radiology, Thomas Jefferson University Hospital, Philadelphia, PA, USA c Department of Neurology, Thomas Jefferson University Hospital, Philadelphia, PA, USA d Department of Neurology, Mt Sinai School of Medicine, New York, NY, USA e Hospital San Vicente de Paul and Medimagen, Medellin, Colombia b

Received 30 December 1999; accepted 8 February 2001

Abstract MRI is a very sensitive imaging modality, however with relatively low specificity. The aim of this work was to determine the potential of image post-processing using 3D-tissue segmentation technique for identification and quantitative characterization of intracranial lesions primarily in the white matter. Forty subjects participated in this study: 28 patients with brain multiple sclerosis (MS), 6 patients with subcortical ischemic vascular dementia (SIVD), and 6 patients with lacunar white matter infarcts (LI). In routine MR imaging these pathologies may be almost indistinguishable. The 3D-tissue segmentation technique used in this study was based on three input MR images (T1, T2-weighted, and proton density). A modified k-Nearest-Neighbor (k-NN) algorithm optimized for maximum computation speed and high quality segmentation was utilized. In MS lesions, two very distinct subsets were classified using this procedure. Based on the results of segmentation one subset probably represent gliosis, and the other edema and demyelination. In SIVD, the segmented images demonstrated homogeneity, which differentiates SIVD from the heterogeneity observed in MS. This homogeneity was in agreement with the general histological findings. The LI changes pathophysiologically from subacute to chronic. The segmented images closely correlated with these changes, showing a central area of necrosis with cyst formation surrounded by an area that appears like reactive gliosis. In the chronic state, the cyst intensity was similar to that of CSF, while in the subacute stage, the peripheral rim was more prominent. Regional brain lesion load were also obtained on one MS patient to demonstrate the potential use of this technique for lesion load measurements. The majority of lesions were identified in the parietal and occipital lobes. The follow-up study showed qualitatively and quantitatively that the calculated MS load increase was associated with brain atrophy represented by an increase in CSF volume as well as decrease in “normal” brain tissue volumes. Importantly, these results were consistent with the patient’s clinical evolution of the disease after a six-month period. In conclusion, these results show there is a potential application for a 3D tissue segmentation technique to characterize white matter lesions with similar intensities on T2-weighted MR images. The proposed methodology warrants further clinical investigation and evaluation in a large patient population. © 2001 Elsevier Science Inc. All rights reserved. Keywords: MRI; Tissue segmentation; white matter disease

1. Introduction In recent years there has been increased effort to identify and characterize, both qualitatively and quantitatively the * Corresponding author. Tel.: ⫹1-215-842-4940; fax: ⫹1-215-8494827. E-mail address: [email protected] (F.B. Mohamed).

intracranial lesions. These efforts have included the introduction of a number of modified MR imaging techniques, such as MR spectroscopy and magnetization transfer imaging [1–7]. In this study we used a fast multispectral 3Dtissue segmentation MR post processing technique for differentiation of several intracranial white matter lesions. Over the years there have been several approaches to tissue segmentation of brain MRI based on one or two sets

0730-725X/01/$ – see front matter © 2001 Elsevier Science Inc. All rights reserved. PII: S 0 7 3 0 - 7 2 5 X ( 0 1 ) 0 0 2 9 1 - 0

208

F. Mohamed et al / Magnetic Resonance Imaging 19 (2001) 207–218

Fig. 1. Multiple sclerosis (a,b,c). Conventional T1, T2 and proton density images of a patient with MS involving the brain white matter. These three MRI images are used as inputs to obtain the 3D segmentation. In the T2-weighted images the MS lesions are seen as bright, homogeneous, high intensity areas located mainly in the periventricular region. (d) A 3D segmented image of the same patient showing two subsets of a lesions, in red and pink. The subset signal represented in red (long arrow), has a high intensity in the T2-weighted spin echo and proton density images, and very low signal intensity in the T1-weighted spin echo image. The other subset, represented in pink (short arrow), has a normal appearing signal intensity on the T1-weighted spin echo image, but has high signal intensity in the T2-weighted fast spin echo and proton density images. The distribution of the red areas is mainly periventricular, although some can also be seen within plaques located further peripherally.

of MR images [8 –10]. They ranged from semi-automatic techniques based on thresholding and k-Nearest Neighbor (k-NN) algorithms to fully automated techniques based on fuzzy connectedness principles. Our earliest works in segmentation [11,12] was based on methodology developed by Cline, et al [13], which used probability and a 2D-feature map incorporating proton density (PD), and T2-weighted MR images as the inputs for tissue segmentation. Later it was demonstrated [14] that by using k-NN segmentation algorithm and inclusion of a third input image could further increase the separation of clusters representing different tissues in the feature space and, thus, produce sharper and clearer representation of these tissues in color-coded segmented images. T1-weighted spin echo (SE) images were used as the third input, because the T1 information would be potentially useful in the characterization of cerebral lesions. An important advantage of this technique is that once we could identify and characterize lesions in the brain as a

specific, “color coded” tissues, we can integrate all these tissues within a volume of interest, (VOI) and measure the lesion volume. In subsequent research studies, we can follow both qualitatively and quantitatively changes in lesions in the brain. Application of this methodology has been validated in studies that demonstrated a correlation between segmentation results of hamster brain tumor model and stereotacitically guided biopsy of brain tumors in humans and histopathology [14]. In-vitro validation of this method using a volumetric phantom (five simulated human tissues), has shown accurate volumetric measurements using the 3D-feature map model (⫾2.5%) [15]. We have also demonstrated the accuracy, inter and intra observer reproducibility [12] of our segmentation technique on normal brain tissues [15]. We hypothesized and proved in earlier studies that T1 contribution introduces substantial new pathophysiological information and produces, greater cluster separation in 3D-feature space leading to better tissue segmenta-

F. Mohamed et al / Magnetic Resonance Imaging 19 (2001) 207–218

209

Fig. 2. Multiple sclerosis (a). A conventional T2-weighted image of a 33-year-old woman with MS. An Axial section at the upper convexity reveals bilateral semiovale. The lesions are scattered in the centrum semiovale. The appearance of these lesions suggests the diagnosis of MS. (b) A 3D-segmented image of the same patient reveals the two subsets of lesions (red, pink) within the plaques as described in the previous MS patient. No significant cortical atrophy is seen.

tion. In this work we have attempted to characterize the intracranial white matter lesions using this 3D-tissue segmentation method. We have selected three white matter lesions which include multiple sclerosis (MS) lesions, subcortical ischemic vascular dementia (SIVD), and lacunar infarctions (LI) lesions in an attempt to better differentiate between these lesions using this 3D tissue segmentation method. These lesions have unique disease characteristics, but produce similar signal characteristics on MR images. All of these lesions are usually visualized as high signal intensity lesions on proton density and T2-weighted spin echo images and have associated low signal intensity in T1-weighted images on the conventional MRI. Hence it is very difficult to identify with MRI, the specific etiology of these three distinct pathological processes. The tissue segmentation procedure may allow the physician(s) to characterize and separate the lesions by identifying multiple tissues present in the lesions that are not readily apparent on routine MRI. The radiologists do not have prior knowledge of the pathology. Along with normal tissues, he or she identifies areas in the brain where pixel intensities suggest that they do not belong there. Only after the post-processing algorithm reconstructs the segmented image, the radiologists are able to clearly visualize the tissue distribution of abnormal areas in the brain, and make the

final diagnosis of the disease. The primary goal of this work was to determine the potential of this image post-processing technique for identification and quantitative characterization of intracranial lesions primarily in the white matter.

2. Methods and materials Forty patients participated in this study: Twenty-eight patients with brain multiple sclerosis, six patients with subcortical ischemic vascular dementia, and six patients with lacunar white matter infarcts. In routine MR images these pathologies were almost indistinguishable. In our tissue segmentation technique, three inputs were utilized. These were T1-weighted spin echo, proton density and T2weighted fast spin echo MRI. In all the patient populations these three sets of MR images were obtained for tissue segmentation. The images were obtained on 1.5T Signa scanner (General Electric Milwaukee, Wisconsin.) Imaging parameters were as follows: two acquisitions and two averages; slice thickness ⫽ 3 mm, interleaved; matrix ⫽ 256 ⫻ 256; field of view (FoV) ⫽ 20 cm; proton density/T2 fast spin echo (FSE): TE1 ⫽ 16 ms, TE2 ⫽ 96 ms; T1 spin echo imaging parameters were: TR ⫽ 500 ms, and TE ⫽ 12 ms. In order to minimize the patient motion a special collar was used between the scans.

210

F. Mohamed et al / Magnetic Resonance Imaging 19 (2001) 207–218

Fig. 3. Subcortical ischemic vascular dementia (SIVD) (a). A conventional T2-weighted image of the same patient. An axial section of the upper convexity reveals homogeneous high intensity lesions involving the white matter of the centrum semiovale. (b) A 3D-segmented image of the upper convexity that reveals one type of homogeneous lesion, represented in red involving a large area of the centrum semiovale. Notice that in contrast to MS, which usually has two subsets of lesions, the white matter lesions of SIVD only has one lesion of homogeneous intensity as visualized in the 3D segmentation. The gray matter in this segmentation image is represented in pink color.

After the scanning the images were transferred to a Sun Sparc Station 10 (Sun Microsystems, Mountain View, California). Initially motion correction algorithm was applied to the images to correct for any misregistration between the MR image series [16]. Next, radiofrequency (RF) inhomogeneities, particularly strong at the edges of RF coil were corrected on all the input images prior to segmentation [17]. We have previously shown that correction of RF nonuniformity improved tissue identification, i.e., improved conspicuity of the lesions on MR images and consequently, contrast resolution of segmented images and improved accuracy of volumetric measurements. Next, in order to reduce the spikes from the MR images, an anisotropic 3Ddiffusion filter [18] was applied to all MR images. This filtering smoothed the tissues without blurring small morphological details and small lesions [19]. A qualified observer (neuroradiologist) “seeded” (select) training tissue samples (40 –50 samples/tissue). The k-Nearest Neighbor (k-NN) segmentation algorithm [20] was used for calculation of the 3D-feature map. This algorithm utilizes the probabilities from the distribution of tissue clusters in the 3D-feature space. Additionally, we performed cluster optimization by searching the densest cluster for each tissue. Only tissue samples from these clusters were used as the input [21]. This greatly reduced operator errors in misclassified seeding process. This data was used to create a stack of color-coded segmented images, and up to 10 tissues were

classified. A connectivity algorithm, [22] along with a dividing cube algorithm [23] was used to construct a surface of selected tissue(s). The detailed methodology of the segmentation procedure used in this study [15] as well as the results of the several pre-processing steps involved in are shown elsewhere [21] and only the final segmentation results are shown in this paper. The segmentation results were reviewed by five qualified observers four of which were board certified (C.F, S.F, F.L, and R.K). They performed the seeding process and/or evaluated the results as well. The observers were blinded during the seeding process, however they were aware of the clinical history during the interpretation of the segmentation results. The actual methodology of the seeding process is also described elsewhere [15].

3. Results 3.1. Lesion characterization 3.1.1. Multiple sclerosis plaques (MS) The cerebral MS lesions were located in the periventricular white matter, and were characterized as bright, homogeneous, high signal intensity areas on fast spin echo and proton density, and T2-weighted MR images collected before segmentation (Figs. 1a– b, 2a). The corresponding areas of low signal intensity were identified in most, but not all, of

F. Mohamed et al / Magnetic Resonance Imaging 19 (2001) 207–218

211

Fig. 4. Lacunar infarct (LI) (a). A conventional T2-weighted image of a 63-year-old man with a history of progressive dementia hypertension and arterosclerosis. An axial section at the midventricular level reveals diffuse cortical brain atrophy and ventriculomegaly. A small high intensity lesion is seen in close proximity to the left periventricular wall (arrow). It is impossible to define whether or not this lesion represents SIVD, multiple sclerosis or a lacunar infarct. (b) A 3D segmented image of the same patient shows that the area visualized in the T2-weighted image is homogeneous and has an intensity close to that of cerebrospinal fluid, represented in blue (arrow). This appearance is most consistent with an old lacunar infarct containing a significant cystic portion.

the T1-weighted spin echo images of the cerebral white matter (1c). Post-segmentation analysis of the 28 MS patients revealed two distinct subsets of signal intensities within the MS plaques: one subset signal is represented in red, and the other in pink (Figs. 1d, 2b). The red subset (Figs. 1d (arrow), 2b), had high signal intensity in the proton density and T2-weighted fast spin echo images, but very low signal intensity in the T1weighted spin echo image, the so-called “black hole”. Therefore, we believe it represents the chronic or more permanent changes within the plaque, noted in conventional MRI studies of patients with secondary progressive MS [24 –27]. The distribution of the red subsets was mainly periventricular, although some red areas could also be seen within the more peripheral subcortical white matter. The red subset was identified in 80% of the secondary progressive MS plaques, and had a different size and configuration than the surrounding subset designated in pink. The pink subset within the MS plaque (Figs. 1d (small arrow), 2b) had high signal intensity in the proton density and T2-weighted fast spin echo images, but isointense T1weighted signal in the images observed before segmentation. Therefore, we believe it represents more acute changes within the plaque, such as demyelination and edema. The distribution of the pink subset was either found surrounding

the red subset or occurred independently in the white matter. The pink subset was identified in all MS patients. 3.1.2. Subcortical ischemic vascular dementia (SIVD) SIVD lesions appear as patchy areas of hyperintensity in the proton density and T2-weighted fast spin echo images of the centrum semiovale and corona radiata (Fig. 3a). Corresponding homogeneous hypointense areas could be seen on the T1-weighted spin echo sequences. However, in contrast to the two distinct subsets of MS lesions observed, distinguished by their different appearances in T1-weighted images, the post-segmentation appearance of SIVD lesions had only one form, represented in red (Fig. 3b). There was no heterogeneity identified in any of the six SIVD patients examined. In several SIVD patients where the histopathological data were obtained, the segmentation-derived images correlated well with the histology, characterized by homogeneous areas of demyelination [28 –31]. The histological data is not presented in this paper. 3.1.3. Lacunar infarcts (LI) The term “lacunar infarct” (LI) was introduced by C. Miller Fisher in 1965 [32,33], and represents a micro infarction ranging from 0.5 to 15 mm in size. These infarctions are caused by occlusion of small perforating branch

212

F. Mohamed et al / Magnetic Resonance Imaging 19 (2001) 207–218

Fig. 5. (a) A conventional T2-weighted image of a 63-year-old patient with a history of stroke and left hemiparesis. The mid-ventricular sagittal section demonstrates a well-defined area of increased intensity in the left basal ganglia involving the anterior limb of the internal capsule. A not so well defined increased intensity area is seen in the opposite side. Diffuse periventricular increased intensity is also visualized. (b) A 3D segmented image of the same patient reveals a mixed lesion located on the right basal ganglia with red and blue sublets (short arrow). This appearance is characteristic of subacute lacunar infarct representing a cystic formation (blue) and an area of gliosis (red). The infarct on the other side has a homogeneous intensity (red) (long arrow).

vessels that may occur in either both the white matter and the central gray matter nuclei of the brain. The histological findings characterizing these lesions are mainly residual cystic changes surrounded by a small area of reactive gliosis [34]. Lacunar infarcts in MR images of six patients were identified as well defined, rounded areas, with high signal intensity in the proton density and T2-weighted fast spin echo images (Figs. 4a, 5a). In contrast, these lesions varied in intensity from isointense to hypointense in the T1weighted spin echo images, depending on their age. Postsegmentation analyses of these lesions correspondingly varied depending upon the age of the lacunae. In older infarctions, the segmented images were characterized by well-defined homogeneous areas of high signal intensity in the proton density and T2-weighted fast spin echo images with cluster values in feature space similar to that of cerebrospinal fluid, corresponding to the cystic changes observed histopathologically. These changes were represented in blue (same as CSF) in the present segmentation analysis (Fig. 4b). However, in subacute infarctions the segmented images were characterized by two different intensities: a perimeter represented in red (Fig. 5b, arrow), and a central core that is identified in blue (Fig. 5b, small arrow). It should be noted that in the image shown in Fig. 4 some

partial volume effects are evident in the gray matter, and this artifact can interfere with the characterization of the pathology. 3.2. Quantitation and follow-up of MS lesions This section of the study involved only MS patients. As we stated above, two statistically distinct subsets within the cerebral MS lesions were readily observed after segmentation. These are shown in Figs. 1 and 2 as red and pink colors. In this investigation for quantitation purposes we summated both subsets into one MS lesion, “colored red”. The total volumetric lesion burden was measured (in cc’s) by integrating all pixels belonging to any given tissue class within the desired VOI in one MS patient [13]. This information was used to determine lesion volume in threedimensional space. To calculate regional lesion burden, we used a modification of the segmentation program that allows us to subdivide the brain into anatomical regions approximating the frontal, parietal, temporal and occipital lobes. The anatomical volumes of the cerebral lobes in this study were based on data obtained from both axial and sagittal MRI sections. Although the borders of the lobes were arbitrarily defined as compared to conventionally recognized anatomic lobar divisions, a consistently good approximation

F. Mohamed et al / Magnetic Resonance Imaging 19 (2001) 207–218 Table 1 Region distribution of MS lesion load in one representative MS patient (cc)

Gray matter White matter CSF MS lesion Vascularity

Frontal

Parietal

Temporal

Occipital

Total volume

196.27 104.78 43.51 27.78 31.72

159.36 101.14 52.67 30.16 15.61

127.47 52.75 31.71 16.61 25.00

88.70 59.39 8.33 9.67 12.34

571.80 318.06 136.22 84.22 84.67

of the volumetric lobar lesion burden was obtained (Table 1). These results are indicative of the potential of 3D tissue segmentation technique for specific applications. A large clinical study with a follow up of large patient population and statistical analysis is beyond the scope of this paper. It is interesting to note that in an individual MR section of the brain, MS lesions often appear as isolated areas (Fig. 6). However, through the extraction of the volumetric reconstruction of all MS lesions, obtained by suppressing the signal from the normal surrounding tissue, confluence of the MS lesions, previously observed only as isolated lesions within individual brain sections, is observed. This takes the form of a “cast” around the ventricle (Fig. 7a, b), providing a unique perspective on the degree of involvement, which confers to the periventricular distribution of the cerebral MS lesion burden. Progression of MS lesion burden was evaluated by comparing serial examinations in three patients. For example, a comparison of the segmentation analysis of one MS patient

213

restudied after a nine-month interval is shown in Figs. 8 and 9, and the volumetric results are shown in Table 2. A significant increase in the cerebral MS lesion burden (⫹17.7%) was identified in this patient with secondary progressive MS. The overall volume of the intracranial compartment did not change. However, the increased MS lesion burden volume was accompanied by a proportionate decrease in cerebral volume of non-affected brain (⫺5.8%), and an increase in CSF volume (⫹18.5%). These results are indicative of the potential of 3D tissue segmentation technique in very specific applications, such as follow up of patients, effect of drug therapy etc.

4. Discussion Although the conventional CT and MRI signal intensities of MS, SIVD and LI lesions can be quite similar, the histopathological substrate found in these conditions have been extensively described [21], and indicate significantly different neuropathological findings in each setting. Due to the different histopathological substrates found in MS, SIVD and LI, we proposed that we would be able to identify and differentiate the lesions of these different conditions by using the MRI post-processing 3D-segmentation technique. In conventional MRI studies of MS, SIVD and LI, these conditions have all been characterized by areas of increased signal intensity in the proton density, and T2-weighted fast spin echo MR images, and low signal intensity in the T1weighted spin echo images [24,32,33]. However, these find-

Fig. 6. Individual MR section showing scattered MS lesions in red color.

214

F. Mohamed et al / Magnetic Resonance Imaging 19 (2001) 207–218

Fig. 7. Axial (b) and sagittal (a) views of extracted MS lesions, from a patient shown in Fig. 6. This volumetric, three-dimensional representation of the total cerebral MS lesion burden, can be displayed and rotated in space to be further visually analyzed. The otherwise unappreciated continuity of multiple lesions, which appear as isolated lesions in individual brain sections, can be seen to form a periventricular “cast”. The volume of the reconstructed image is rapidly quantified.

ings are not considered to be specific for any of these disorders, and careful correlation with the clinical images is always advised. In addition, the MRI features found in MS lesions in particular have been extensively studied with many modifications of the imaging technique, to enhance the characterization of the lesions. These have included serial MRI studies to follow the natural history of the disease. Magnetization transfer imaging (MTI), and more recently MR spectroscopy, have also been applied to the characterization of MS lesions [1–7], to better define the nature of the tissue disruption in this disorder. The 3D MRI post-processing segmentation technique based on a 3D-feature map (PD, T2 and T1) can discriminate, and thus identify, different tissue subsets within the MS plaques in clinically definite secondary progressive MS. The segmentation technique is not programmed to separate data into “distinct” clusters. It is programmed to identify each voxel as belonging to one group of the tissue. Therefore if identification is unclear (statistically poor) the voxels with different colors will not create distinct areas, but will present themselves in a randomized fashion, i.e., spatially mixed colors. These findings corroborate, and are thus validated, by previous MRI studies demonstrating heterogeneity within the same MS plaque identified by using MTI. In the MTI studies, lower magnetization transfer ratios, corresponding to the red areas in our studies, were felt to represent chronic changes, while higher magnetization transfer ratios, corresponding to the pink areas, were felt to represent more acute changes, such as edema or demyelination. The high signal intensity subset in the proton density and T2-weighted fast spin echo images with a low signal intensity in the T1-weighted spin echo images, represented in red (Fig. 1a, b), best correlates with the distribution of T1weighted spin echo images obtained before segmentation in conventional MRI studies. For this reason, we believe that

this subset represents more permanent tissue changes, such as gliosis, as has previously been reported [24 –27]. In contrast, the subset with a high signal intensity in the proton density, T2-weighted fast spin echo and isointensity in the T1-weighted spin echo images, represented in pink (Fig. 1d, 2b), best correlates with the localization of the T2-weighted fast spin echo images obtained before segmentation in conventional MRI studies. Therefore, we believe that this subset most likely represents edema and demyelination. The recognition of these lesion subsets, and quantification of their relative proportions present in a given patient may provide an important clue as to why a patient fails to respond to a particular form of therapy administered to alter the course of their disease. This information may also serve as an adjunct inclusion criteria in therapeutic clinical trials of specific agents selected on the basis of their capacity to either reduce inflammation, edema and demyelination on one hand, or to resolve gliosis and promote remyelination on the other, rather than simply selecting patients on the basis of their clinical course alone. In SIVD, also known in the literature by a variety of terms such as subcortical arteriosclerotic encephalopathy, leuko-raiosis and Binswanger’s disease, [35] the MRI finding observed on routine T2-weighted fast spin echo images are almost indistinguishable from those found in MS. In routine clinical practice, the differential diagnosis between these two diseases is usually best made based upon the age of the patient and the clinical presentation. Therefore, older patients with MS present a diagnostic problem, and the post-segmentation findings observations are very helpful in distinguishing the lesions of MS and SIVD. In MS, there is heterogeneity of the plaque observed in post-segmentation, while in SIVD, we studied there is homogeneity of the high signal intensity T2-weighted fast spin echo image (Fig. 3b), which corroborates the histopathological images of demyelination [25,27].

F. Mohamed et al / Magnetic Resonance Imaging 19 (2001) 207–218

215

Fig. 8. a. Proton density, b. T2-weighted, c. T1-weighted, and d. 3D-segmented image of a patient with secondary progressive MS. Notice the two subsets (red and pink areas) of lesions previously described.

The third group of patients studied had lacunar infarcts identified in the white matter. These lesions can all be confused with MS plaques or SIVD changes on routine MRI studies. The histopathology of these infarcts demonstrates an evolution from a subacute to a chronic stage. The postsegmentation images from subacute and chronic infarcts closely correlate with the histopathological changes encountered. There is a central area of necrosis with cyst formation in the older lesions, surrounded by an area of reactive gliosis in the more subacute lesions [24 –26]. In the segmented images of older lacunar infarcts the signal intensity inside the cystic area is very similar to that of CSF (Fig. 4b), represented in blue, while in subacute infarctions, there is

also a peripheral rim of reactive gliosis indicated in red (Fig. 5b). It is important to note that some partial volume effects may be present in the gray matter, and this artifact can interfere with the characterization of the pathology as consisting of red and blue (similar to CSF) colors. The preliminary results of this 3D MRI post-processing segmentation strongly suggests that it can be used as a reliable method for differentiating between the cerebral white matter lesions of MS, SIVD, and LI, which has importance not only diagnostically, but providing treatment decisions as well. With this technique it is possible to separate these different lesions based upon their cluster location in multifeature space. This is shown in Table 1, and

216

F. Mohamed et al / Magnetic Resonance Imaging 19 (2001) 207–218

Fig. 9. a. Proton density, b. T2-weighted, c. T1-weighted, and d. 3D-segmented image of the same patient shown in Fig. 8, using the same parameters, as studied nine months later in the clinical course. Notice the increase in cerebral MS lesion burden during this time interval. Quantitative measures of the component tissues (brain, CSF, cerebral MS lesion burden, and total volume) are provided in Table 2.

demonstrated in the figures qualitatively. To the best of our knowledge we are not aware of any published work which has attempted to characterize these lesions as segmentation

Table 2 Change in MS load (cc) with time in one patient Tissue

Start

9-months

% change

Brain CSF MS Total volume

786.1 145.0 79.9 1011

740.8 171.8 94.1 1007

⫺5.8 ⫹18.5 ⫹17.7 0

does, and in this work discriminate between three similarly appearing lesions. These results are indicative of the potential of 3D tissue segmentation techniques for specific applications. A clinical study with a follow up of large patient population and statistical analysis is outside the scope of this paper. In addition to improved identification and lesion characteristics, the presented segmentation technique can also be utilized for measuring the volume of the lesions identified, as well as any of the other components of the brain (i.e., total CSF volume, white matter volume, gray matter volume, etc). In serial studies of cerebral in particular, it is thus,

F. Mohamed et al / Magnetic Resonance Imaging 19 (2001) 207–218

potentially, possible to utilize this technique to compare the change in lesion burden over time, consistency with the natural history of the disorder, and to assess the impact of treatment on lesion burden accumulation. Recently, we developed fast multispectral segmentation based on a 4D-feature map [38]. In this technique we have demonstrated its superiority as compared with that based on 3D-feature map (p ⬍ 0.001). This new technique can be utilized to further evaluate application of segmentation to identify, quantify and follow-up of intracranial lesions. Correlation of these cerebral MRI findings with neuropsychological testing of the patient in each of these disorders could lead to a better understanding of the clinical significance of the brain tissue changes observed in MS, SIVD and LI, as was demonstrated by findings in MS [35–37]. Because this technique also provides a tool with which to follow changes in the lesions over time, it also has the potential to play a significant role in the evaluation of therapies in the treatment of these diseases.

[12]

[13]

[14]

[15]

[16] [17]

[18] [19]

References [1] Mehta RC, Pike GB, Enzmann DR. Improved detection of enhancing and non-enhancing lesions of multiple sclerosis with magnetization transfer. Amer Journal Neuroradiology 1995;16:1771– 8. [2] Filippi M, Rovaris M. Magnetisation transfer imaging in multiple sclerosis. [Review] [41 refs] Journal of Neurovirology 2000;6(Suppl 2):S115–20. [3] Filippi M, Iannucci G, Cercignani M, Assunta RM, Pratesi A, Comi G. A quantitative study of water diffusion in multiple sclerosis lesions and normal-appearing white matter using echo-planar imaging. Archives of Neurology 2000;57(7):1017–21. [4] Narayana PA, Wolinksy JJ, Jackson EF, McCarthy M. Proton MR spectroscopy of gadolinium-enhanced multiple sclerosis plaques. J Magnetic Resonance Imaging 1992;2:263–70. [5] Hielhle JF, Lenkinski RE, Grossman RI, et al. Correlation of spectroscopy and magnetization transfer imaging in the evaluation of demyelinating lesions and normal appearing white matter in multiple sclerosis. Magnetic Resonance Medicine 1994;32:285–93. [6] Hiehle JF, Grossman RI, Ramer KN, Gonzalez-Scarno F, Cohen JA. Magnetization transfer effects in MR-detected multiple sclerosis lesions: comparison with gadolinium-spin-echo images and non-enhanced T1-weighted images. American J Neuroradiology 1994;16: 69 –77. [7] Loevner LA, Grossman RI, McGowan JC, Ramer KN, Cohen LA. Characterization of multiple sclerosis plaques with T1-weighted MR and quantitative magnetization transfer. American J Neuroradiology 1995;16:1473–9. [8] Clark MC, Hall LO, Goldgof DB, Velthuizen R, Murtagh FR, Silbiger MS. Automatic tumor segmentation using knowledge-based techniques. IEEE Transactions on Medical Imaging 1998;17(2):187– 201. [9] Clarke LP, Velthuizen R, Phuphanich S, Schellenberg JA, Arrington M, Silbiger M. MRI stability of three supervised segmentation techniques. Magnetic Resonance Imaging 1993;11:95–106. [10] Udupa JK, Wei L, Samarasekera S, Miki Y, van Buchem MA, Grossman RI. Multiple sclerosis lesion quantification using fuzzyconnectedness principles. IEEE Transactions on Medical Imaging 1997;16(5):598 – 609. [11] Vinitski S, Seshagiri S, Mohamed FB, et al. Tissue characterization by MR: data segmentation using 3D-feature map. In: Vernazza G,

[20] [21]

[22]

[23]

[24]

[25]

[26]

[27]

[28]

[29]

[30]

[31]

217

Venetsanopoulos AN, Braccini C (Eds), Image Processing Theory and Applications, Amsterdam: Elsevier Science Publishers B.V. 1993;325– 8. Vinitski S, Gonzalez C, Burnett C, et al. Tissue segmentation by high resolution MRI: improved accuracy and stability. Proc IEEE Eng Med Biology 1994;16:577– 8. Cline HE, Lorensen WE, Kikinis R, Jolesz F. Three-dimensional segmentation of MR images of the head using probability and connectivity. J Computer Assist Tomography 1990;14:1037– 42. Vinitski S, Gonzalez CF, Andrews D, Knobler R, Curtis M, Mohamed F, Gordon J, Khalili K. In vivo validation of tissue segmentation based on a 3D feature map using both a hamster brain tumor model and stereotacitically guided biopsy of brain tumors in man. J Magnetic Resonance Imaging 1998;8:814 –9. Vinitski S, Gonzalez C, Mohamed FB, et al. Improved intracranial lesion characterization by tissue segmentation based on a 3D-feature map. Magnetic Resonance Medicine 1997;37:457– 69. Nissanov J, Madi S, Vinitski S. Distance-based subset alignment of MR images. Radiology 1997;205(p):51. Mohamed FB, Vinitski S, Gonzalez CF, Faro S, Burnett C, Ortega HV, Iwanaga T. Image non-uniformity correction in high field (1.5T) MRI. Proc IEEE Eng Med Biology 1995;17:36 –7. Gerig G, et al. Nonlinear anisotropic filtering of MRI data. IEEE Transaction on Medical Imaging Vol. II. No 2, 1992. Perona P, Malik J. Scale space and edge detection using anisotropic diffusion. Proc IEEE Workshop on Computer Vision. Miami, FL 1987;6 –22. Cover TM, Hart PE: Classification IEEE transaction on information theory. Nearest Neighborhood Pattern Vol. 13, 1967;21–7. Mohamed FB, Vinitski S, Faro S, Gonzalez CF, Khalili K, Lublin F, Ortega HV. Optimization of tissue segmentation based on multispectral 3D feature maps. Magnetic Resonance Imaging 1999;17:403–9. Cline HE, Dumoulin CL, Hart Jr HR, et al. 3D reconstruction of the brain from MRI using a connectivity algorithm. Magnetic Resonance Imaging 1987;5:345–9. Cline HE, Lorensen WE, Ludke S, Crawford CR, Teeter BC. Two algorithms for the three-dimensional reconstruction of tomograms. Medical Physics 1988;15:320 –7. Amspach JP, Gounot D, Rumbach L, Chambron J. In vivo determination of multiexponential T2 relaxation in the brain of patients with multiple sclerosis. Magnetic Resonance Medicine 1991;9:107–13. Lacomic D, Osbakken M, Gross G. Spin-lattice relaxation (T1) times of cerebral white matter in multiple sclerosis. Magnetic Resonance Medicine 1986;3:194 –202. Larsson HBW, Frederiksen J, Petersen J, et al. Assessment of demyelination edema and gliosis by in vivo determination of T1 and T2 in the brain of patients with acute attack of multiple sclerosis. Magnetic Resonance Medicine 1989;11:337– 48. van Waderveen MA, Barkhoff F, Hommes OR, Polman CH, Frequin STFM, Valk J. T1-SE more specific than T2-SE in identifying disabling lesions in multiple sclerosis. A quantitative follow-up study. Proceedings of the Society of Magnetic Resonance SMR, 1994. Scarpeli M, Salvolini V, Diamanti L, Montironi R, Chiromoni L, Maricotti M. MRI and pathological examination of post mortem brains: the problem of white matter high signal areas. Neuroradiology 1994;36:393– 8. Tien RD, Felsberg GJ, Ferris NJ, Osumi AK. The dementia’s: correlation of clinical features, pathophysiological and neuroradiology. American J Radiology 1993;161:245–55. Braffman B, Zimmerman RA, Trojanowski JQ, Gonzalez NK, Hickey WF, Scalaepfer WW, Brain MR. Pathologic correlation with gross and histopathology. Hyperintense white matter foci in the elderly. American J Neuroradiology 1988;9:629 –36. Fazekas F, Kleinert R, Offenbacher H, et al. The morphologic correlate of incidental punctuate white mater hyperintensities on MR images. Amer Journal Neuroradiology 1991;12:915–21.

218

F. Mohamed et al / Magnetic Resonance Imaging 19 (2001) 207–218

[32] Fisher CM. Lacunes: small, deep cerebral infarcts. Neurology 1965; 15:774 – 84. [33] Fisher CM. Lacunar strokes and infarcts: A review. Neurology 1982; 32:871– 6. [34] Braffman B, Zimmerman RA, Trojanowski JQ, Gonatas NK, Hickey WF, Schalaepfer WW, Brain MR. Pathologic correlation with gross and histopathology. Lacunar infarction and Virchow-Robin spaces. Amer Journal Neuroradiology 1988;9:621– 8. [35] Adams JH, Corsellis J, Ducher LW. “Greenfields’s Neuropathology” John Wiley & Sons, New York, 1984. [36] Nesbit GM, Forbes GS, Scheithauer BW, Okazaki H, Rodregues M. Multiple sclerosis: histopathologic and MR and/or CT correlation in

37 cases at biopsy and three cases at autopsy. Radiology 1991;180: 467–74. [37] Gonzalez CF, Mitchell DR, Swirskey-Sacchetti T, et al. Correlation between structural brain lesions and emotional and cognitive function in patients with multiple sclerosis: An MRI study. Neuroradiology 1991;(suppl)123–124. [38] Gonzalez CF, Swirskey-Sacchetti T, Mitchell DR, et al. Distributional patterns of MS lesions: An anatomical and clinical correlation based on MRI findings. J Neuroimaging 1994;4:188 –95. [39] Vinitski S, Iwanaga T, Gonzalez C, Knobler R, Andrews D. Fast tissue segmentation based on a 4D feature map in characterization of intracranial lesions. J Magnetic Resonance Imaging 1999;9:768.