Sulci density map to aid in use of apparent diffusion coefficient for therapy evaluation

Transcription

1 Available online at Magnetic Resonance Imaging 26 (2008) Sulci density map to aid in use of apparent diffusion coefficient for therapy evaluation Lars A. Ewell4, Christopher J. Watchman, Kurt Wharton Department of Radiation Oncology, University of Arizona, Tucson, AZ , USA Received 27 September 2006; revised 16 February 2007; accepted 10 April 2007 Abstract In order to more accurately asses variations in the apparent diffusion coefficient used for therapy evaluation, we have studied the variation in sulci density in the human brain. Sagittal, axial and coronal magnetic resonance imaging scans have been analyzed to determine the change of the coefficient of variance of pixel intensity as a function of position. In the sagittal direction, relative to the 50% most medial slices, we find an 11.0%F4.8% (S.D.) decrease in the next 25% (12.5% on each side) of the slices. The most lateral 25% of the slices had less of a decrease and more variation: 7.0%F12.2%. Similar variations were observed in axial and coronal scans. D 2008 Elsevier Inc. All rights reserved. Keywords: Sulci; Density; Variation; Diffusion; Coefficient; Noise 1. Introduction Using diffusion weighted magnetic resonance imaging (DWMRI), it has recently become possible to calculate an apparent diffusion coefficient (ADC) for a region of interest (ROI). A rise in the ADC of water has been linked with effective cancer therapy and vice versa [1,2]. Due in part to a relative lack of motion artifacts, this technique is often employed in the brain. However, the brain has unique features that can possibly complicate the interpretation of a variation in the ADC. In particular, the density of sulci has the potential to impact how significant a change in the ADC of a ROI may be. The sulci are the fissures in the brain that are filled with cerebral spinal fluid (CSF). Since CSF is a free fluid, an ROI that contains a high density of sulci should, in principal, result in a higher ADC than an ROI that has a low density of sulci. Some of the large sulci in the human brain, such as the lacrimal, are named and vary little in size or position from patient to patient. However, for the smaller sulci, there is significant intra- and interpatient variation and a means by which to quantify the sulci density is therefore an important consideration. Using the coefficient of variance of pixel intensity (COVPI) as a metric for sulci density, we 4 Corresponding author. address: lewell@ .arizona.edu (L.A. Ewell). have analyzed sagittal, axial and coronal magnetic resonance imaging (MRI) scans of a number of patients. Significant interpatient, as well as intrapatient variations were observed. The trends of the variations, however, between different patients are similar Motivation In the field of radiation oncology, the ADC has been investigated in an attempt to determine an early indication of therapy efficacy. It is hoped that information of this nature can aid in clinical decisions and enable earlier use of second line therapies. Recent measurements of ADC changes have predicted disease progression and overall survival [3]. Prior to radiation oncology, water mobility and DWMRI were used in an attempt to better diagnose cerebral ischemia [4]. ADCs and DWMRI are still used widely in the study of ischemia. It is known that partial volume averaging of CSF and parenchyma introduces significant bias into ADC calculation/measurements [5]. Furthermore, it has been reported that by using CSF-suppressed DWMRI scans [fluid-attenuated inversion recovery (FLAIR)], more accurate and predictive values of ADC can be calculated [6].Itis assumed that a similar effect can be expected when using ADCs to evaluate cancer therapy efficacy. A major difference between ischemia and neoplasia is the introduction of a different type of tissue into the brain, that of the tumor. In addition, the resection cavity and resulting edema X/$ see front matter D 2008 Elsevier Inc. All rights reserved. doi: /j.mri

2 L.A. Ewell et al. / Magnetic Resonance Imaging 26 (2008) Fig. 2. Signal-to-noise ratio vs. sagittal position. Fig. 1. SMPTE test pattern. (A) Rectangle with high (0.97) value of COVPI. (B) Rectangle with an intermediate value (0.65) of COVPI. (C) Rectangle with low value (0) of COVPI. associated with standard tumor treatment further complicates diagnosis. With this in mind, we have analyzed MRI scans in a number of patients in an attempt to quantify the spatial density and spatial variation of sulci in the brain. 2. Methods and materials 2.1. Sulci density quantification In a typical T 1 -weighted image, due to its long T 1 time (4500ms [7]), CSF is heavily suppressed relative to gray and white matter. In view of this fact, we have investigated the standard deviation of pixel intensity as a means by which to quantify the density of the CSF filled sulci in the brain. The standard deviation, r, is the square root of the variance, which is the average of the squares of the deviations from the mean, i.e., 2!# r ¼ 1 X n 1=2 4 ðx k lþ 2 ð1þ n k¼1 this is a test pattern used by the Society of Motion Picture and Television Engineers (SMPTE). In Fig. 1, this test pattern is displayed with pixel size and 8-bit gray scale resolution. In Rectangle A, an ROI has been chosen that includes interspersed bright and dark stripes. These strips have either the minimum intensity, 0 (dark), or full intensity, 255 (bright), reflecting the 8 bit resolution. If an equal number of dark and bright pixels are included in the ROI, l =127.5 and r =127.5 so that the COV= 1.0. This is close to the measured value of COV=0.97. This is to be compared to a region of high sulci density where MRI signal differences between the where x k is the intensity of the kth pixel, l is the average pixel intensity and n is the number of pixels in the ROI. The coefficient of variance (COV) is then defined as the ratio of the standard deviation to the average, or COV ¼ r l : ð2þ It can be viewed as a normalized standard deviation for an ROI. As suggested by the American College of Radiology [8], a useful tool for quantification, such as Fig. 3. (A) Medial ROI. (B) Lateral ROI.

3 22 L.A. Ewell et al. / Magnetic Resonance Imaging 26 (2008) Table 1 Scan characteristics Scan set Scan type Slice thickness (mm) Slice spacing (mm) Number of slices Field strength (T) Total width (mm) 1 T 1 IRP 3DSPGR T T 1 FLAIR T 1 SE IRP 3DSPGR, Inversion Recovery Prepared 3 Dimensional Spoiled Gradient Echo; SE, Spin Echo. CSF-filled sulci, and normal white matter would be high. However, since the pixel intensity range in a T 1 -weighted MRI scan is less than in this test pattern, we expect the COV for an actual MRI scan to be less. For example, if the pixel intensity of the bright stripe is 200 and for a dark stripe, is 100, then we have that l =150 and r =50 so that the COV=1/3. Similarly, if the volume of the sulci is large enough to occupy half of the pixels in an ROI with like pixel range, we can expect a corresponding value for the COV. For comparison, a mixed ROI (B) has a value of COV= 0.65, while a uniform ROI (C) has a value of COV= Data acquisition In order to apply a similar method to MRI scans from different patients, a method by which to estimate the signalto-noise ratio (SNR) is needed. Using a previously described procedure [9], the SNR is determined via SNR ¼ 0:655 l ð3þ r air where again, l is the mean signal intensity in the ROI, r air is the standard deviation of the dark background (air), and the factor accounts for the fact that the (gaussian) noise in the raw data is centered about zero. In Fig. 2, the SNR for four different sets of sagittal magnetic resonance scans is plotted as a function of the fractional distance. As can be seen, all four exhibit a minimum at a medial point, where the space between the hemispheres is filled with CSF. In Fig. 3, a medial and lateral slice from a sagittal T 1 weighted scan set are shown. To obtain the data, the cerebrum only was contoured such that the skull, corpus callosum, ventricles or other brain anatomy is avoided. To account for variation, the background (air) noise was subtracted from the signal in the ROI to obtain the noise adjusted value, r NA. That is, r NA ¼ r ROI r air ð4þ with an average noise level of 6% (F1% S.D.). The COVPI is then obtained in the ROI as described above. The software ImageJ [10] was utilized in this image analysis. In Table 1, some of the scan characteristics of the some of the data are shown. 3. Results/plots The COVPI as a function of the slice number is plotted to estimate how the sulci density varies in a sagittal direction across the cerebrum. In Fig. 4, these data are plotted. As can be seen in this figure, the COVPI starts out at a relatively high value in the extreme left periphery of the cerebrum (low slice number), then goes through a minimum near the mid left hemisphere, and then to a maximum in the medial slices near the corpus callosum. Rough symmetry is observed between the left and right hemispheres of the brain. For reference, the slices shown in Fig. 4. Covariance vs. slice number for sagittal scan. Fig. 5. COVPI rebinned and averaged.

4 L.A. Ewell et al. / Magnetic Resonance Imaging 26 (2008) Fig. 7. Axial COVPI. Fig. 6. Interpatient covariance comparison. Fig. 3A and B correspond to slice numbers 47 and 14 in this figure, respectively. In order to more easily compare COVPI slice distribution between patients, it is useful to rebin and average the data from different slices. To this end, the middle 50% of the slices are grouped together; then the next, 25% (12.5% on the left and right) and finally, the most lateral 25%. These rebinned data are plotted in Fig. 5 (note: the scale in this figure is expanded relative to the same data plotted in Fig. 4). As can be seen in this plot, the data form a bw,q reflecting high sulci density near the lateral peripheries, followed by a minimum in mid hemisphere, and then a maxima in the most medial slices near the corpus callosum. Display of the data in this form facilitates comparison between different patients. In Fig. 6, four different data sets are displayed, along with the average. As can be seen in the figure, all four of the data sets exhibit the same trend of a decline in the COVPI when going from the most medial slices outward. In three of the four sets, there is a rise going from the slices in the mid hemisphere, to the most peripheral slices. As indicated in the table, the top two data sets were taken on a 3T machine (GE, Signa Excite), while the sets with a lower COVPI were taken on a 1.5T machine (Marconi Medical, Eclipse). In Table 2, the mean data are displayed. As can be seen in this table, the medial deviation is more uniform in the mid hemisphere region, with a lower standard deviation (5.0%) and smaller range than the peripheral/ lateral regions. In addition to the sagittal direction, analogous data has been analyzed in a similar fashion in the axial and coronal directions. In Figs. 7 and 8, data for axial and coronal directions are plotted respectively. In Fig. 7, both T 1 and T 2 weighted images are plotted for a number of different patients. As can be seen, a general trend of higher COVPI for more superior slices is seen. In the coronal direction, an increase in COVPI is seen in going from a posterior to an anterior location. 4. Discussion 4.1. Trends Combining data from three different directions, several characteristics emerge from the trends: (1) all else being equal, an inferior, posterior point in the mid hemisphere can be expected to have the lowest COVPI; (2) a medial, anterior, superior point can be expected to have the highest; (3) variation in the sagittal direction is greatest, due to an increase near the lateral and medial peripheries of the individual hemispheres. As indicated above, these data were analyzed to more accurately interpret ADCs used for therapy evaluation. However, these data were derived from mostly T 1 and to a lesser extent, T 2 - weighted images. Although the resolution and image intensity achieved in DWMRI scans is in general less than for non diffusion-weighted images, we expect that these same general trends will hold. Studies are underway, whereby these characteristics will be studied in a larger cohort of patients longitudinally Estimated effects Previously, ischemic researchers have estimated that a voxel containing 20% of CSF results in an ADC that is Table 2 Mean rebinned data left/right average Location Medial difference Standard deviation Range Periphery 7.0% 12.2% 23.1% (Patient 3) 4.4% (Patient 2) Mid hemisphere 11.0% 4.8% 17.7% (Patient 1) 7.6% (Patient 2)

5 24 L.A. Ewell et al. / Magnetic Resonance Imaging 26 (2008) % higher than for normal brain parenchyma [6]. Assuming a similar rise, an estimate can be made regarding how much influence glioma location can have on an ADC due to the difference in sulci density. The placement of the ROI used in determining an ADC is somewhat subjective. Often, a T 1 weighted, gadolinium-enhanced MRI is used by a radiologist or radiation oncologist to contour the tumor. The region between the tumor periphery and the associated edema is frequently difficult to discern. To estimate the effect of sulci density on an ADC, several assumptions will be made: (1) the central core tumor does not cross any sulci; (2) the ROI used in determining the ADC includes an amount of edema which does include sulci: the ROI of sulci infiltration (ROISI); (3) the tumor is spherical in shape with a 5-cm diameter; (4) the edema surrounding the tumor is also spherical in shape, with a shell thickness of 2 cm, 1 cm of which is included in the ROI. This is depicted graphically in Fig. 9. The tumor volume V T =4/3pr 3 = 4/3p(2.5 cm) 3 =65.4 cm 3. The ROI volume V ROI =4/3p(3.5 cm) 3 =179.6 cm 3. The volume of the ROI in which sulci infiltrate is V ROISI =4/3p[(3.5 cm) 3 (2.5 cm) 3 ]=114.2 cm 3. Due to the spherical shape of the ROI, although the sulci infiltration shell (ROISI) is only 1 cm thick (out of a total radius of 3.5 cm), it constitutes over 50% (63.6%) of the total volume. If this ROI is located, e.g., in a medial region of the cerebrum, we can expect that the sulci density may be approximately 11% higher than if it were located in a region of the mid hemisphere, according to Table 2. If we further assume that due to this difference, there is an 11% higher volume of CSF in the ROISI, then there may be an approximately 9% [ (CSF factor [6])100 =8.8%] rise in the ADC due to partial volume averaging of CSF and tumor, regardless of treatment efficacy. Since attempts are often made to minimize the amount of normal tissue treated, the thickness of the ROISI shell may be lower than the 1 cm referred to above. This would result in value lower than 9%. On the other hand, it has recently been pointed out that gliomas often metastasize across sulci via CSF [11], which could result in sulci infiltration of the core tumor (Fig. 9), resulting in a value higher than 9%. In addition to the absolute location of the tumor, migration also should be considered. Recently, it has been established that glioma cells can migrate at speeds of up to 100 l/h in animal models [12]. If a tumor migrates from a low-density sulci region, to a high one, the value of an ADC for a ROI can be expected to rise due to partial volume averaging of CSF and tumor, again regardless of treatment efficacy. 5. Conclusion Fig. 9. ROI with partial sulci infiltration. We have argued that the COVPI in MRI scans is a useful measure of sulci density. By analyzing a number of scans, we have demonstrated that this metric shows substantial variation throughout the brain. These variations have implications for the use of apparent diffusion coefficients in therapeutic evaluation. In particular, tumor location in the cerebrum can have a substantial effect on an ADC calculation, due to volume averaging of CSF and tumor. Although there are significant differences between patients, the interpatient trends are similar. By studying these trends, we have made an estimate of the difference in ADC values between a high sulci density region and a low one. Additional prospective data will improve predictive capabilities. Acknowledgment The authors would like to acknowledge helpful discussion with Baldassarre Stea of the Department of Radiation Oncology at the University of Arizona Medical Center. References Fig. 8. Coronal COVPI. [1] Ross BD, Moffat BA, Lawrence TS, Mukherj SK, Gebarski SS, Quint DJ, et al. Evaluation of cancer therapy using diffusion magnetic resonance imaging. Mol Cancer Ther 2003;2: [2] Theilmann R, Borders R, Trouard TP, Xia G, Outwater E, Ranger- Moore J, et al. Changes in water mobility measured by diffusion MRI

6 L.A. Ewell et al. / Magnetic Resonance Imaging 26 (2008) response of metastatic breast cancer to chemotherapy. Neoplasia 2004;6: [3] Hamstra D, Chenevert TL, Moffat BA, Johnson TD, Meyer CR, Mukherj SK, et al. Evaluation of the functional diffusion map as an early biomarker of time-to-progression and overall survival in highgrade glioma. PNAS 2005;102: [4] Warach S, Chien D, Li W, Ronthal M, Edelman RR. Fast magnetic resonance diffusion-weighted imaging of acute human stroke. Neurology 1992;42: [5] Liu G, Van Gelderen P, Duyn J, Moonen CT. Single-shot diffusion MRI of human brain on a conventional clinical instrument. Magn Reson Med 1996;35: [6] Latour L, Warach S. Cerebral spinal fluid contamination of the measurement of the apparent diffusion coefficient of water in acute stroke. Magn Reson Med 2002;48: [7] Haacke E, Brown R, Thompson M, Venkatesan R. Magnetic resonance imaging physical principals and sequence design. New York7 John Wiley and Sons; [8] dacr standard for teleradiologyt, revised 2002, effective 1/1/03. [9] Firbank M, Coulthard A, Harrison R, Williams E. A comparison of two methods for measuring the signal to noise ratio on MR images. Phys Med Biol 1999;44:N [10] See [11] Fujimura M, Kumabe T, Jokura H, Shirane R, Yoshimoto T, Tominaga T. Intractable vomiting as an early clinical symptom of cerebrospinal fluid seeding to the fourth ventricle in patients with high-grade astrocytoma. J Neurooncol 2004;66: [12] Farin A, Suzuki S, Weiker M, Goldman J, Bruce J, Canoll P. Transplanted glioma cells migrate and proliferate on host brain vasculature: a dynamic analysis. Glia 2006;53: