Brain MR Image Normalization in Texture Analysis of Multiple Sclerosis

Transcription

1 Brain MR Image Normalization in Texture Analysis of Multiple Sclerosis C.P. Loizou, Member, IEEE, IEEE, M. Pantziaris, I. Seimenis, C.S. Pattichis, Senior Member Abstract A problem that occurs in texture analysis and quantitative analysis of magnetic resonance imaging (MRI), is that the extracted results are not comparable between consecutive or repeated scans or, within the same scan, between different anatomic regions. The reason is that there are intra-scan and inter-scan image intensity variations due to the MRI instrumentation. Therefore, image intensity normalization methods should be applied to magnetic resonance (MR) images prior to further image analysis. The objective of this work was to investigate six different MRI intensity normalization methods and propose the most appropriate for the pre-processing of brain T2-weighted MR images acquired from 22 symptomatic untreated multiple sclerosis (MS) subjects and 10 healthy volunteers. Following image normalization, texture analysis was carried out in original and normalized images for normal appearing white matter (NAMW) and MS lesions, detected in transverse T2- weighted MR images. The best normalization method (Histogram Normalization (HN)) demonstrated a smaller Kullback Leibler divergence (0.05, 0.06) suggesting appropriateness for pre-processing MR images used in texture analysis of MS brain lesions. This is a prerequisite step in the assessment of texture features as surrogate markers of disease progression. Index Terms MRI, multiple sclerosis, intensity normalization. M I. INTRODUCTION ultiple Sclerosis (MS) is a chronic idiopathic disease that results in multiple areas of inflammatory demyelination within the central nervous system. Within individuals the clinical manifestations are unpredictable, particularly with regard to the development of disability [1]. Correlations between MS and disability were investigated in [2], whereas in [3] disease subgroups were classified based on their MS disease severity. Texture features quantify macroscopic lesions and characterize macroscopic changes that may be undetectable using conventional measures of lesion volume and number [1]. Manuscript received July 4, 2009, accepted 29 August C.P. Loizou, is with the Department of Computer Science, School of Sciences, Intercollege, P.O.Box 51604, CY-3507, Limassol, Cyprus (phone: ; loizou.c@ lim.intercollege.ac.cy; loizou.christos@ucy.ac.c. M. Pantziaris is with the Cyprus Institute of Neurology and Genetics, Nicosia, Cyprus; ( pantzari@cing.ac.c. I. Seimenis, is with the Medical Diagnostic Center Ayios Therissos, 2033 Nicosia, Cyprus ( yseimen@phys.uoa.gr). C.S. Pattichis, is with the Department of Computer Science, University of Cyprus, Nicosia, Cyprus; ( pattichi@ucy.ac.c. Texture analysis, therefore is used widely in magnetic resonance imaging (MRI) enabling disease characterization and quantification of disease distribution. It is worth noting that these techniques may provide information that is not visible to human eye. However, texture parameters are sensitive to the acquisition conditions including magnetic resonance (MR) protocols, MR scanners and MR adjustments [4]. This yields unwanted intensity variations which may affect the results of image analysis. An intensity normalization correction stage is therefore essential. It was shown in [4] and [5] that normalization methods improve image compatibility by reducing the variability introduced by different gain settings, different operators, different equipment, and facilitates MR image comparability. Different approaches to the above problem have been proposed in the literature. In [6] the influence of normalization was studied for different acquisition protocols. Meier et al. [7] proposed an intra- and an interscan normalization method for processing serial MRI scans for direct quantitative evaluation. In [8] a histogram mapping was proposed, whereas in [9] a contrast stretch normalization based on the maximum and minimum gray scale values in the image was introduced. In [10], a normalization method for brain MRI using histogram evenorder derivative analysis was proposed, whereas in [11], a histogram matching method to correct the variations due to scanner sensitivity was presented. Finally in [5], a normalization method was proposed, where the original image histogram is stretched and shifted in order to cover all the available gray levels of the image. The normalization methods investigated in our study aimed to a much simpler methodology compare to those proposed in the literature. The objective of our study was to implement and investigate six different normalization methods for brain MR images and propose the most appropriate one for the texture analysis of MS brain lesions. II. MATERIALS AND METHODS A. Study group and MRI acquisition Twenty two subjects (11 males, and 11 females), aged 30.7±11.5 years (mean age ± standard deviation), with a clinical isolated syndrome (CIS) of MS and MRI detectable brain lesions were scanned twice with an interval of 6-12 months. All subjects were untreated and clinically examined by an experienced neurologist following the MRI and at the end of the study were given an EDSS (expanded disability

2 status scale) score [12]. Additionally, the brain of 10 healthy, age-matched (30.8±7.6 years) volunteers (4 males, and 6 females) was scanned for image texture analysis on healthy brain normal white matter (NWM). Five of those volunteers were MRI-scanned in a second imaging session, a few days after the first one, to allow for the study of interscan image intensity normalization. The images used for analysis were obtained using a T2-weighted turbo spin echo pulse sequence (TR=4408ms, TE=100ms, echo spacing=10.8ms). The MRI protocol and the acquisition parameters are given in detail in [12]. All detected brain lesions were manually segmented by an experienced MS neurologist and confirmed by a radiologist. B. Image intensity normalization methods The following MRI intra-scan intensity normalization methods were implemented: 1) Contrast Stretch and Normalization (CSN) [9]: The global maximum (g max ) and minimum (g min ) of the original image are firstly estimated by excluding the image background which is normally 0. A 9x9 pixel window is used in the neighborhood of g max and g min to compute the g hithres and g low-thres respectively by averaging all pixels in the corresponding windows. That is values above g max are equated to g hi-thres and values below g min are equated to g lowthres. 2) Intensity Scaling (IS): The neurologist selects homogeneous high intensity regions of interest (cerebrospinal fluid or High Intensity Region-g HIR ) and low intensity regions (air from the sinuses or Low Intensity Region-g LIR ) from the brain. The image intensities are mapped to the values between g HIR and g LIR. This was implemented using the Matlab function gscale. 3) Histogram Stretching (HS): The following normalization or contrast stretching transformation for increasing the dynamic range of the resulting image was carried out: g( gmin f ( = (1) gmax gmin where g( is the original image gray scale value at coordinates x and y, g max and g min represent the maximum and the minimum gray scale values in the original image respectively, and f( the resulted normalized pixel. 4) Histogram Normalization (HN) [5]: The original image histogram is stretched, and shifted in order to cover all the gray scale levels in the image as follows: ghir glir f ( =.( g( g ) + g (2) min LIR g g max min If the original histogram of the initial image g(, starts at g min and extends up to g max brightness levels, then we can scale up the image so that the pixels in the new image, f(, lie between a minimum level (g LIR ) and a maximum level (g HIR ). This is done by scaling up the intensity levels according to (2). 5) Gaussian Kernel Normalization (GKN) [13]: A local normalization algorithm is used that normalizes the local mean and variance of the image estimated by a Gaussian Kernel using smoothing operators. The resulted image is as follows: g( m f ( f ( = (3) σ f ( where m f ( and σ f ( represent the mean and standard deviation of the original image respectively. 6) Histogram Equalization (HE) [8]: Image histogram equalization was carried out using the Matlab function (a) Original MR image (c) Original MRI histogram (b) Normalized MR image (d) Normalized MRI histogram Fig. 1. (a) MR original image of the brain from a patient with MS, (b) the normalized image (with the method HN [5]), (c) its original histogram (median (IQR): 72 (11)) and (d) the normalized histogram (median (IQR): 81 (13)). Fig. 2. Box plots for the Kullback Leibler distance for the 6 different normalization methods for 0 and 6-12 months for the 22 MS subjects studied. IQR values are shown above the box plots. In each plot we display the median, lower, and upper quartiles and confidence interval around the median. Straight lines connect the nearest observations with 1.5 of the Inter-Quartile Range (IQR) of the lower and upper quartiles. Unfilled rectangles indicate possible outliers with values beyond the ends of the 1.5 x IQR.

3 histeq. C. Intrascan and interscan ROI manual segmentation For intrascan normalization purposes, NWM, as well as high and low intensity regions were segmented by an experienced radiologist in two distantly apart slices from the same scan of the healthy volunteers. For interscan normalization purposes, similar regions were segmented in corresponding transverse slices obtained from two different imaging sessions. In the patient group, all detectable brain lesions were identified and segmented by an experienced neurologist and confirmed by a radiologist, along with areas of high and low signal intensity. D. Texture analysis Texture features (Statistical Features and Spatial Gray Level Dependence Matrices as proposed by Haralick et al. (mean, median, standard deviation, contrast, entrop [14], and shape parameters ( X coordinate maximum length, Y coordinate maximum length, area, perimeter, perimeter 2 /area, eccentricity, equivalence diameter, major axis length, minor axis length, centroid, convex area, and orientation) were extracted from all MS lesions detected and segmented [14], [15]. E. Evaluation metrics 1) Distance Measures In order to identify the most discriminant features and cross-evaluate between the different normalization methods the following distance measure was computed for each feature [16]: 2 2 diszc = mza mzs / σ za + σ (4) zs where z is the feature inde c if o indicates the original image set and if f indicates the normalized image set, m za and m zs are the mean values and σ za and σ zs are the standard deviations of the original and normalized classes respectively. The most discriminant features are considered to be the ones with the higher distance values [16]. 2) Kullback Leibler Divergence Instead of measuring the similarity, we measure the dissimilarity of two histograms H 1 and H 2 by the Kullback- TABLE I KULLBACK LEIBLER DIVERGENCE (KLD) DISTANCE BETWEEN DIFFERENT SLICES FROM SAME SCAN (INTRASCAN) AND BETWEEN CORRESPONDING SLICES FROM DIFFERENT SCANS (INTERSCAN) OF NORMAL VOLUNTEERS FOR NON NORMALIZED (ORIGINAL) IMAGES AND ALL THE SIX DIFFERENT NORMALIZATION METHODS INVESTIGATED (MEAN±SD) Method Intrascan Interscan Orig ± ±0.135 CSN 0.156± ±0.015 IS 0.112± ±0.017 HS 0.178± ±0.023 HN 0.099± ±0.010 GKN 0.165± ±0.010 HE 0.185± ±0.006 Leibler divergence (KLD) or relative entropy between two distributions [17], given by: Porig ( i) KL( Iorig ( i), I norm( i)) = Porig ( i) log (5) i Pnorm( i) where, P orig and P norm are probability distributions corresponding to I orig and I norm, respectively, and i represents the gray level index. The KL divergence yields a positive value. A higher score indicates a higher dissimilarity, hence a higher possibility of a histogram or a feature shift [17]. 3) Univariate Statistical Analysis The Wilcoxon rank sum test was used in order to detect if for each texture feature a significant difference (S) or not (NS) exists between the original and the normalized images at p<0.05. We also calculated the coefficient of variation, CV%, which describes the difference as a percentage of the pooled mean value with CV%=(σ/m)*100 [5]. Furthermore, the correlation coefficient, ρ, between the original and the normalized features was calculated. III. RESULTS For the healthy subjects studied, the KLD between different slices from the same scan was smaller when the images were normalized with the HN method as shown in Table I. The same method was also proved to be the most efficient, in terms of reduced dissimilarity, when the KLD of normalized images obtained from two imaging sessions was compared to that of the corresponding original images (Table I). The Wilcoxon rank sum test performed for the KLD between the original and HN-normalized data showed a significant difference for both the intra-scan and inter-scan normalization processes. Figure 1(a) shows an original MR image of a healthy volunteer and Fig. 1(b) the normalized image using the HN method. The original and normalized histograms are shown in Fig. 1(c) and Fig. 1(d) respectively. The median (IQR) values for the original and normalized images were 72 (11) versus 81 (13) respectively. Table II presents texture features at 0 and 6-12 months extracted from the original (Orig.) and normalized MR images (CSN, IS, HS, HN, GKN, HE). It is shown that the HN method exhibits the smallest correlation coefficient (ρ=0.14), and a large distance (dis=0.31). Furthermore, the median value does not significantly change after normalization. The Wilcoxon rank sum test performed on the texture features between the original and the normalized images for 0 and 6-12 months, presented in Table III, showed that in most of the cases the normalization process resulted in significant differences (S) between the original and the normalized images. Figure 2 presents box plots for the KLD for 0 and 6-12 months for all normalization methods investigated for the 22 MS subjects. It is clearly illustrated that the HN method exhibits the smallest KLD at both 0 and 6 months (median

4 values of 0.05 and 0.06 respectivel which means that this method brings about the smallest divergence to the original histogram. The Wilcoxon test performed for differences in the KLD at 0 and 6-12 months between the original and the normalized images, showed that no significant differences (NS) were found between the original and the normalized images (p values for the 6 different normalization methods were: CSN:0.73, IS:0.29, HS:0.31, HN:0.19, GKN:0.43, HE:0.81). IV. DISCUSSION The objective of this study was to implement and investigate six different image normalization methods and propose the most suitable for analysis of brain MR images. We have shown in this paper that the HN method for MRI normalization proposed in [5] (see Table I and Fig. 2), in which the original histogram is stretched and shifted in order to cover a wider dynamic range, yields KLD values which are lower than the corresponding values given by the other methods tested herein. Therefore, it could be argued that it is an appropriate method for MRI normalization. This method is based on g HIR and g LIR intensity measurements. Homogeneous high intensity regions of interest (cerebrospinal fluid or High Intensity Region-g HIR ) and low intensity regions (air from the sinuses or Low Intensity Region-g LIR ) from the brain are selected, which are then used in (2) for scaling up the initial intensity levels of the image. A method also proposed in [6] (see Fig. 2, IS) demonstrated similar results with a KLD median (IQR) values of 0.06 (0.12) and 0.07 (0.16). Table III shows, that the median, standard deviation, and contrast of brain MS lesions are smaller after normalization for all different normalization methods. The smaller standard deviation and median of the sampling distribution after normalization shows that this reflects the true population parameters more accurately. Similar findings (i.e. smaller mean, median and standard deviation after normalization) were also found in [18], where a statistical method of correcting spatially dependent image pixel intensity non-uniformity based on differences in local tissue intensity distributions was proposed. It was also shown in this study, that the methods IS and HN exhibit higher distance values (d=0.38 and d=0.31) when compared to the rest of the methods (see Table II). The proposed normalization method allows the scanner sensitivity variations and variations due to repeatability studies to be largely corrected and thereby facilitating meaningful comparisons between MRI data sets obtained at different times and/or different subjects. By normalizing the histogram of the whole brain we introduced an automatic procedure with little sensitivity to pathological or morphological changes between the different image data sets. The method does not depend on knowledge of the scanner calibration and thus can be used on retrospective data. Several other studies with more complicated algorithms and very promising results have been proposed in the literature for MRI normalization that are briefly discussed below. In [10] the utility of even order derivative analysis in MRI histogram normalisation has been demonstrated. It was shown that good white matter peak discrimination was achieved even when significant overlap existed between gray matter and white matter peaks, as this is the case with the T2-weighted brain images. Furthermore, the ability of TABLE II MEDIAN VALUES OF TEXTURE FEATURES AND STATISTICAL ANALYSIS FOR ALL LESIONS AT 0/6-12 MONTHS FOR THE DIFFERENT NORMALIZATION METHODS INVESTIGATED Method Median S/NS STD S/NS Contr S/NS Entr S/NS CV(%) S/NS ρ dis S/NS Orig. 125/116 S(0.007) 16/15 NS(0.11) 95/78 S(0.03) 5.1/5.1 NS(0.9) 13.2/13.3 NS(0.68) CSN 114/106 S(0.04) 22/21 NS(0.24) 122/97 NS(0.1) 5.7/5.7 NS(0.8) 19.7/20.2 NS(0.88) S(0.007) IS 117/107 S(0.02) 20/19 NS(0.22) 87/77 NS(0.2) 5.7/5.6 NS(0.8) 17.3/17.9 NS(0.51) NS(0.09) HS 115/106 S(0.03) 22/20 NS(0.26) 102/89 NS(0.1) 5.7/5.7 NS(0.8) 18.7/19.2 NS(0.71) S(0.003) HN 116/107 NS(0.07) 20/19 NS(0.52) 116/10 NS(0.4) 5.7/5.7 NS(0.8) 17.4/18.2 NS(0.54) NS(0.12) GKN 116/112 NS(0.1) 19/20 NS(0.41) 109/11 NS(0.8) 5.7/5.7 NS(0.7) 16.8/18.1 NS(0.13) S(0.03) HE 115/111 S(0.003) 10/15 NS(0.09) 56/59 NS(0.22) 5.3/5.4 NS(0.3) 6.9/7.4 S(0.05) S(0.003) STD: Standard Deviation, Contr.: Contrast, Entr.: Entropy, CV: Coefficient of Variation, ρ: Correlations coefficient, S/NS: Significantly (S) different at p<=0.05, non-significantly different (NS) at p>0.05, the p value given in parentheses TABLE III WILCOXON RANK SUM TEST PERFORMED ON THE TEXTURE FEATURES EXTRACTED FROM ALL LESIONS BETWEEN THE ORIGINAL AND THE PROCESSED IMAGES AT 0 AND 6-12 MONTHS Method Median STD Contr Entr CV% ρ Median STD Contr Entr CV% ρ Original vs Normalization Methods at 0 months Original vs Normalization Methods at 6-12 months CSN S S S S S S S S S S S S IS S S S S S S S S NS S S S HS S S S S S S S S S S S S HN S S S S S S S S NS S S S GKN S S NS S S S NS S S S S S HE S NS S S S S S NS S S S S STD: Standard Deviation, Contr.: Contrast, Entr.: Entropy, CV: Coefficient of Variation, ρ: Correlations coefficient, S/NS: Significantly (S) different at p<=0.05, non-significantly different (NS) at p>0.05, the p value given in parentheses

5 the normalization procedure to correct the global intensity variations over time was demonstrated by the high degree of reproducibility of an automatic brain segmentation algorithm following intensity normalization. In another study [7], an image post processing method for integrating multiple serial MRI scans into a volume for direct quantitative evaluation of the temporal intensity profiles was proposed. A significant error reduction was observed when applying tissue specific inter-scan intensity normalization, suggesting that intensity variations above 3% can be reliably detected. A histogram matching method was proposed in [11] for correcting the variations in scanner sensitivity due to differences in scanner performance. It was shown that the method reduced the variation in white matter intensities from 7.5% to 2.5% and provided a method to remove the threshold dependency in lesion volume measurement with global thresholding in patients with MS. Furthermore, in [6], the influence of MRI acquisition protocols, and gray level normalization methods on texture classification were evaluated. The results showed that the normalization method and the acquisition protocol influence the effectiveness of the classification. More specifically, if no normalization was applied, the classification errors depend on the MR acquisition protocols. However, when using normalization, no relationship was observed and the classification results were significantly improved. MRI analysis has become a powerful tool in the diagnosis of brain disease [1]-[4]. Pixel intensity variations between the same and consecutive MRI scans i.e. intra-scan and inter-scan variations, complicate the method of quantitative MRI analysis [6]. Improvements in the measurement and pre-processing of the image may therefore have a significant impact in the clinical diagnosis, image analysis, and computer aided diagnosis. The simple method of histogram intensity normalization proposed in [5] and tested in this study can help in this direction, however more studies with larger datasets are required. This will enable an accurate computation of texture features that may provide information for better and earlier differentiation between normal tissue and MS lesions and in assessing disease progression. [2] M. Filippi, D.W. Paty, L. Kappos, F. Barkhof, D.A. et al., Correlations between changes in disability and T2-weighted brain MRI activity in multiple sclerosis: a follow-up study, Neur., vol. 45, pp , [3] J. Dehmeshki G.J. Barker, and P.S. Tofts, Classifications of disease subgroups and correlation with disease severity using magnetic resonance imaging whole-brain histograms: Application to magnetisation transfer ratios and multiple sclerosis, IEEE Trans. Med. Imag., vol. 21, no. 4, pp , [4] A. Simmons, P.S. Tofts, G.J. Barkers, S.R. Arridge, Sources of inhomogeneity in spin echo images at 1.5 T, Magn. Reson. Med., vol. 32, pp , [5] M. Nixon, A. Aguado, Feature Extraction & Image Processing, Newnes, [6] G. Collewet, M. Strzelecki, and F. Marriette, Influence of MRI acquisition protocols and image intensity normalization methods on texture classification, Magn. Reson. Imag., vol. 22, pp , [7] D.S. Meier, R.G. Guttman, Time-series analysis of MRI patterns in multible slcerosis, Neuroimage, vol. 20, pp , [8] R.C. Gonzalez, R.E. Woods, Digital Image processing, 2nd Ed., Prentice Hall, [9] S. Koptenko, Contrast Stretch and Normalization, Matlab FEX 2006, [10] J.D. Christensen, Normalization of brain magnetic resonance images using histogram even-order derivative analysis, Mag. Reson. Imag., vol. 21, pp , [11] L. Wang, H.-M. Lai, G.J. Barker, D.H. Miller, P.S. Tofts, Correction for variations in MRI scanner sensitivity in brain studies with histogram matching, Magn. Res. Medic., vol. 39, no. 2, pp , [12] A.J. Thompson, and J.C Hobart, Multiple sclerosis: assessment of disability and disability scales, J. Neurol., vol. 245, no. 4, pp , [13] G. Heush, F. Gardinau Lighting normalisation algorithms for face verification, IDIAP research institute, Internal report, IDIAP-com-03, pp.1-40, March [14] R.M. Haralick, K. Shanmugam, and I. Dinstein, Texture Features for Image Classification, IEEE Trans. Systems, Man., and Cybernetics, vol. SMC-3, pp , Nov [15] C.P. Loizou, C.S. Pattichis, I. Seimenis, E. Eracleous, C.N. Schizas, and M. Pantziaris, Quantitative Analysis of Brain White Matter Lesions in Multiple Sclerosis Subjects: Preliminary Findings, IEEE Proc. of the the 5th Int. Conf. Inf. Techn. And Appl. In Biomed., ITAB 2008, Shenzhen, China, May 30-31, 2008, pp [16] C. Christodoulou, C. Pattichis, M. Pantziaris, and A. Nicolaides, Texture-Based Classification of Atherosclerotic Carotid Plaques, IEEE Trans. Med. Imag., vol. 22, no. 7, pp , [17] J.P.W. Pluim, J.B.A.Maintz, M.A.Viergever, f-information measures in medical image registration, IEEE Trans. Med. Imag., vol. 23, no. 12, pp , Dec [18] C. DeCarli, D.G.M. Murphy, D. Teichberg, G. Campbell, G.S. Sobering, Local histogram correction of MRI spatially dependent image pixel intensity nonuniformity, J. Magn. Res. Imag., vol. 6, no. 3, pp , ACKNOWLEDGMENT This work was funded through the project Quantitative and Qualitative Analysis of MRI Brain Images ΤΠΕ/ΟΡΙΖΟ/0308(ΒΙΕ)/15, 12/ /2010, of the Program for Research and Technological Development , of the Research Promotion Foundation of Cyprus. REFERENCES [1] F. Fazekas, F. Barkof, M. Filippi, et al, The contribution of magnetic resonance imaging to the diagnosis of multiple sclerosis, Neur., vol. 53, pp , 1999.