- Email: [email protected]
A novel method based on independent component analysis for brain MR image tissue classification into CSF, WM and GM for atrophy detection in Alzheimer’s disease
Biomedical Signal Processing and Control 40 (2018) 41–48
Contents lists available at ScienceDirect
Biomedical Signal Processing and Control journal homepage: www.elsevier.com/locate/bspc
Research paper
A novel method based on independent component analysis for brain MR image tissue classification into CSF, WM and GM for atrophy detection in Alzheimer’s disease Rupali S. Kamathe a,∗ , Kalyani R. Joshi b a b
Department of Electronics and Telecommunication, College of Engineering, Pune, Maharashtra, India Department of Electronics and Telecommunication, PES’s MCOE, Pune, Maharashtra, India
a r t i c l e
i n f o
Article history: Received 4 April 2017 Received in revised form 26 July 2017 Accepted 5 September 2017 Keywords: Independent component analysis Band expansion process Brain MRI Support vector machines Alzheimer’s disease
a b s t r a c t Brain Magnetic Resonance Image (MRI) plays a vital role in diagnosis of diseases like Brain Tumor, Alzheimer, Multiple Sclerosis, Schizophrenia and other White Matter Lesions. In most of the cases accurate segmentation of Brain MRI into tissue types like Cerebro-Spinal Fluid (CSF), White Matter (WM) and Grey Matter (GM) is of interest. The diagnostic accuracy of expert and non-expert Radiologists can be improved with accurate and automated tissue segmentation and classification system. Such system can also be used for trainees to understand the individual tissue distribution in MRI scans. In this paper, we propose a novel automated tissue segmentation and classification method based on Independent Component Analysis (ICA) with Band Expansion Process (BEP) and Support Vector Machine (SVM) classifier which with input as T1, T2 and Proton Density (PD) scans of patient, provides output as CSF, WM and GM indicating the possible atrophy in brain which can help in diagnosis of Alzheimer’s disease (AD). The objective of this work is to test the effectiveness of ICA with different input images generated using BEP for accurate brain tissue segmentation by validating results with different quality metrics. The novel method for generating input images for ICA has been implemented and segmented tissues are used for atrophy detection. The BEP + ICA + Thresholding + ‘SVM trained with Grey Level Co-occurrence Matrix (GLCM) based texture features’ is giving 100% tissue classification accuracy for test samples under consideration. © 2017 Elsevier Ltd. All rights reserved.
1. Introduction Structural MRI is one of the important neuroimaging modality with high-resolution imaging and high brain tissue contrast capabilities. It is particularly used to look for brain tumors, stokes, blood clots, or other abnormalities that might account for Multiple Sclerosis (MS) or Alzheimer’s. MRI scans are acquired by different pulse sequences specified by three MR tissue parameters: spin–lattice (T1), spin–spin (T2) relaxation times, and Proton Density (PD) [5]. These are labeled as T1 weighted, T2 weighted, and PD weighted (Fig. 1). The constituents of brain, such as Gray Matter (GM), Cerebrospinal Fluid (CSF),White Matter (WM), Glial Matter, Fat, Muscle/Skin, etc. show unique characteristics under a magnetic field. However, the major tissue types of brain are CSF, GM and
∗ Corresponding author. E-mail addresses: [email protected] (R.S. Kamathe), [email protected] (K.R. Joshi). http://dx.doi.org/10.1016/j.bspc.2017.09.005 1746-8094/© 2017 Elsevier Ltd. All rights reserved.
WM; which have been distributed in T1, T2 and PD. As a result, spatial as well as tissue characteristics based features can be extracted from these MRI scans [6]. Reader based classification methods for these tissues are non-reproducible, and are practically difficult for the large amounts of data. Thus, development of fully automatic and accurate brain tissue classification from MRI in case of various disease symptoms like Tumors, MS, AD and other White Matter Lesions (WML) is of great interest [13,17] and is a challenging task. AD, the most common type of dementia is a major cause of disability worldwide. It can be detected at an early stage with the help of MRI so as to avoid irreversible damage of the brain with proper treatment plan. MRI can depict ‘atrophy’ − a decrease or shrinkage in the size of different areas of the brain caused due to wasting away of brain tissues in response to a disease process [10] like AD. Fig. 2 shows the various stages of Alzheimer’s disease. As can be seen in Fig. 2, MRI shows increasing atrophy as the disease progresses from MCI to AD (with no atrophy for healthy/normal controls). Measuring atrophy based on tissue volume distribution is important to diagnose, to monitor the disease
42
R.S. Kamathe, K.R. Joshi / Biomedical Signal Processing and Control 40 (2018) 41–48
Fig. 1. MRI Axial Scans.
Fig. 2. Normal Control (NC), Mild Cognitive impairment (MCI) and AD (T1 Weighted Axial Brain MR Images).
progression and in turn helps in the planning of treatment. Most of the segmentation methods are sensitive to the tissue intensity overlaps caused by limitations of MR image acquisition process (partial volume effect). There are three software packages widely used in neuroimaging community for structural and functional brain imaging study: FMRIB Software Library (FSL) [14], Statistical Parametric Mapping (SPM) [15] and BrainSuite [12]. FSL provides very effective tools like BET, FAST, FIRST, FLIRT etc. for structural MRI scans like T1-w or T2-w. FAST needs brain extracted images as input, which can be obtained using BET. SPM needs MATLAB and is based on a mixture model clustering algorithm. Kazmi K et al. in [18] presented detailed performance evaluation of these 3 methods. Chris et al. in [13] described and validated a completely automatic, non-parametric brain tissue classification system, based on a probabilistic anatomical atlas (model) for different tissue labels. Also, in [11] comparative study of different model based tissue classification methods is presented. Among other methods, ICA is also very popular [5,6,8,16]; often as a pre-processing step of SVM based classifier [6,16] and uses different types of structural MRI scans such as T1-w, T2-w and PD as an input. An application of ICA to MR image (MRI) analysis was investigated by Nakai et al. in [8] for contrast enhancement of GM and WM, till then it was in wide use for functional MRI (fMRI) [4]. In paper [8], study revealed that the involvement of grey or white matter in brain tumor cases and demyelination in the case of MS were enhanced and visualized in IC images. Also, in the same study [8], the potential of ICA for further analysis of anatomical images by enhancing the contrast among normal tissue or between normal and pathologic tissues has been investigated. In paper [6], Ouyang et al. pointed out two problems which were not addressed by Nakai et al.:i. Implementing Over Complete ICA (OC ICA) − number of image pulse sequences used for acquisition is generally smaller than the number of brain substances of interest and ii. ICA algorithm is initialized with initial random projection vectors to generate Independent Components (ICs). Because of this ICs generated are random and hence image evaluation cannot be performed until all ICs are generated. In [6], in order to mitigate this issue, the OC-ICA is used in conjunction with a feature extractionbased classifiers like- SVM and Fisher’s linear discriminant analysis (FLDA) for tissue classification. The performance of the method/algorithm for Tissue Segmentation can be validated by comparing the segmented tissue with Ground Truth image. For AD, for evaluating the disease progression, individual tissue volumes need to be measured and atrophy
should be measured with some parameter/s. Sadek in [10] has suggested three measures − Atrophy Ratio (AT), Alzheimer disease factor (ADF) and Progressive AD rate; for this purpose. In this paper, we present the automatic brain tissue classification based on ICA and SVM with novel method for generating input images (using BEP) for ICA to overcome limitations of ICA (mentioned in Section 2). The set of input images is selected with combination of original images (T1, T2 and PD) and band generated images (using different BEP); to generate independent Components (ICs). The approach is towards generating ICs which are as independent as possible with separate IC for each of the tissue of interest such as WM, GM and CSF. The idea for achieving accurate tissue classification is based on training of SVM with texture based spatial features evaluated on the most accurately segmented tissue samples (of WM, GM and CSF). Four different segmentation methods (Model I–IV) are proposed. The performance of all is tested and compared with different quality metrics like Mutual Information (MI), Tanimoto Index (TI) [16], Similarity Index (SI), Precision and Recall [20]. The paper also describes a case study for Alzheimer’s disease. In this study, the segmented tissues for AD cases are used further for calculation of Atrophy Ratio (AT). This paper is organized as follows: Section 2 introduces the ICA and its limitations. Section 3 is about BEP to create more inputs for ICA. Section 4 is about SVM classifier and GLCM Features. Section 5 describes the database used for experimentation. Section 6 gives the detailed methodology implemented in this work. Section 7 describes the various quality parameters used for quantitative analysis of methods implemented and results obtained. Section 8 is for results and discussions. Section 9 is about case study of AD followed by Conclusion in section 10. 2. Independent component analysis (ICA) ICA [1–4] is a linear transformation method. If ‘x’ denotes an m-dimensional random variable; the problem is then to find a function ‘f’ so that the n-dimensional transform s = (s1; s2; . . .. . .. sn) T defined by s = f (x) has some desirable properties and can be represented s = Wx where W is a matrix to be determined. The problems associated with ICA, of ICA being an overdetermined system with the number of samples (M) is usually greater than the sources to be separated (S) is pointed out in [6]. There were generally no solutions for such case. On the other hand, for M < S, ICA becomes an under-determined system. Nakai et al. [8] assumed that M ≥ S, where M in an MR imaging system consists of T1, T2, PD images [6,8] and their unique combinations. In [6] S is considered as the number of brain tissue substances, which includes water, blood, fat, GM, WM, CSF, and muscle, as signal sources to be separated; thus M < S and the ICA must deal with an over-complete representation of a mixed model, in which case many solutions are possible. In this work, Fast ICA: efficient and popular algorithm for ICA invented by Aapo Hyvärinen at Helsinki University of Technology [3,4] has been used. The classical application of the ICA model is that of blind source separation [1,7]; that is, no prior information about distribution of different tissues of interest in input images is required to be known. 3. Band expansion process (BEP) Ouyang et al. in [6] resolved the issue of having less number of input (band) images for ICA in MRI analysis by introducing the Band Expansion Process (BEP). The idea of BEP is to capture the correlation between original MR band images (T1, T2 and PD) and generate a set of second-order statistical band images [5]. The second-order statistics that can be used for BEP includes autocorrelation, cross-
R.S. Kamathe, K.R. Joshi / Biomedical Signal Processing and Control 40 (2018) 41–48
correlation, and nonlinear correlation [17] to create nonlinearly correlated images. In this work, we used fusion rules based on Principal Component Analysis (PCA) & Discrete Wavelet Transform (DWT) [9] for generating band images and found that the methods works well. Ouyang et al. pointed out in [6] that in most of the cases the band images obtained using cross correlation are generally sufficient to implement ICA. We also confirm the same; but the combination of band images obtained using cross correlation, PCA or DWT based methods with original T1, T2 and PD scans, as an input to ICA gives the best segmented tissues of interest. Our approach of generating band images using fusion based methods (as shown in Fig. 3) is inspired by the problem pointed out in [5] for auto correlated band images. Using auto correlated band images may sometimes cause nonsingularity problems in matrix computation because they are self-correlated and usually very close to the original images [5]. Fig. 3(a) and (b) respectively shows the PCA- based and DWT- based fusion scheme implemented for generation of band images. In Fig. 3(a), P1 and P2 are Principal Components of Images I1 and I2 respectively, which are combined by the rule P1I1 + P2I2. If I1is T1sequence, then I2 can be T2 or PD sequence. Likewise, different combinations of T1, T2 and PD sequences can be selected for both the methods. In DWT based fusion scheme, first order DWT is used. Then each of the sub-band pixels in LL, HL, LH and HH bands are combined using Pixel averaging to generate Fused Wavelet Coefficients. Inverse DWT is implemented to obtain the band image.
4. Support vector machine (SVM) and GLCM features SVM [22,24]: It often happens that the sets to discriminate are not linearly separable in finite dimensional space and hence there is need to map this space into a much higher-dimensional space, assuming that the separation will be easier in that space. A SVM [24] constructs a hyper plane or set of hyper planes in a high dimensional space, which can be used for classification. A good separation is achieved by the hyper plane that has the largest distance to the nearest training data points of any class (called functional margin). The SVM computes the simple dot product of the variables in the original space and may employ kernel tricks such as Gaussian Radial Basis Function, polynomial (homogeneous/inhomogeneous), Hyperbolic tangent etc. In this paper, multi-class SVM is implemented with 3 class labels as CSF, WM and GM. GLCM features: For feature extraction, method proposed by Haralick [23], namely, the Spatial Gray-Level Dependence method (SGLDM) (The corresponding matrix is labeled as GLCM- Gray Level Co-occurrence Matrix) is used. For each of the CSF, WM and GM image with N number of grey levels, the GLCM is a square matrix G of order N, where the (i, j)th entry of G represents the number of occasions a pixel with intensity i is adjacent to a pixel with intensity j. The normalized cooccurrence matrix is obtained by dividing each element of G by the total number of co-occurrence pairs in G. The adjacency can be defined to take place in each of the four directions (horizontal−0◦ , vertical−90◦ , left−135◦ and right diagonal−45◦ ). The Haralick’s texture features (Contrast, Correlation, Energy, and Homogeneity) are calculated for each of these directions of adjacency. These features are selected such that they show maximum similarity within the class and minimum similarity between the classes; so as to have good separation between classes under consideration (CSF, WM and GM).
43
5. Data set used for experimentation MRI Data set for Normal cases: McGill University [19]: Simulated Images, 0% noise and 0% INU, Stereotaxic Co-Ordinate System-(0.70,18) and Image slice Thickness–1 mm,3 mm,5 mm,7 mm and 9mm- 3 scans (T1, T2 and PD) of each case. Number of cases is 5. Thus 15 number of images available for a particular slice thickness (say 1 mm) comprising T1, T2 and PD for single experiment. Axial scans are selected as shown in Fig. 2. MRI Data set for AD cases: Harvard Medical School [21]: Real Patients’ data, Number of AD cases: 5; Axial scans are selected. No further pre-processing is required on selected T1, T2 and PD images. Training data for SVM: Four Haralick’s features as mentioned in Section 4 are extracted on each of the tissues classes- CSF, WM and GM obtained by segmentation of 15 number of input images (as described in section 5 (i)). Thus 60 numbers of training features are available for each of the tissue types- CSF, WM and GM respectively. 6. Quantitative analysis Mutual Information (MI) Metric [25]: used for BEP It gives the amount of information transferred from input image to Band Image. It is based on the measure of the degree of dependence of the two random variables A and B (any of the two input images T1, T2 and PD). Considering two input images A, B and F a new fused/band image, the amount of information that F contains about A and B can be calculated using MI. Following measures are used to compare, A: Extracted/segmented tissue and B: Ground truth of that tissue (shown in Fig. 5); to know how well the tissues of interest are segmented. Tanimoto index (TI) [16]: TI = A∩ B/A∪ B Given two sets A and B, if they are completely different, then TI = 0. TI=1 if the sets are completely the same (Tissue obtained after segmentation is exactly same as that of the ground truth). Similarity Index (SI) [20]: SI = |A∩ B|/| (A + B) /2| Precision [20]: Precision = |A∩ B|/|B| Precision gives how much amount of extracted tissue is precise to ground truth image. Recall: Recall = |A∩ B|/|A| Atrophy Ratio (AT) [10]: AT = (GM + WM) / (GM + WM + CSF) AT small values indicate highly atrophied brain. 7. Methodology PART A: Implementation of segmentation methods to obtain tissues of interest (CSF, WM, GM) Fig. 4 shows the four methods (named Model I–IV) implemented for segmentation of MRI scans (T1, T2 and PD) into CSF, WM and GM. For this experimentation, dataset as mentioned in Section 5 is used.
44
R.S. Kamathe, K.R. Joshi / Biomedical Signal Processing and Control 40 (2018) 41–48
Fig. 3. PCA and DWT Based Image Fusion as Band Expansion Process [9].
Fig. 4. Four Methods for Tissue segmentation.
Fig. 5. Automatic Tissue classification System for PART B.
Model I is based on simple Histogram based thresholding for segmentation of input MR Images into CSF, WM and GM. For Model II and Model III, input is not limited to T1, T2 and PD, but more number of input images is generated using band Expansion Processes (BEP), which includes images generated by performing − Autocorrelation, Cross correlation, PCA based Fusion and DWT based fusion on T1, T2 and PD MR Scans for experimentation. In Model II, thresholding is applied on these band images to generate CSF, WM, and GM. In Model III, 3 band images and 3 original images are given as an input to ICA to overcome the problem of OC-ICA. The ICs generated by Model III are further processed for skull removal and then,
thresholding is applied to obtain 3 tissues of interest- CSF, WM and GM. In Model IV, only 3 original MRIs- T1, T2 and PD are used as input to ICA to obtain 3 ICs. These ICs are processed further for skull removal and then, thresholding is applied to obtain3 ICs asCSF, WM and GM. PART B: Implementing Multi-Class SVM for automatic classification of segmented tissues into- CSF, WM and GM (Fig. 5) Tissue images (CSF, WM, GM) obtained using each of the above models (I–IV) described in PART A, are compared with Reference/Ground Truth images (shown in Fig. 6) with four parametersTI, SI, Precision and Recall. Table 2 shows the detail results for 1 mm slice thickness images. Same experimentation is repeated for 3 mm, 5 mm, 7 mm and 9 mm images. Based on the values of these met-
R.S. Kamathe, K.R. Joshi / Biomedical Signal Processing and Control 40 (2018) 41–48
45
Fig. 9. Tissue Segmentation results obtained using Model III (ICA+ Skull Removal +Thresholding) as described in PART A.
Fig. 6. Reference/Ground Truth Images [19].
Fig. 10. Tissue Segmentation results obtained using Model IV (BEP+ICA+Skull Removal +Thresholding) as described in PART A.
Table 1 MI Quality Metric.
Fig. 7. Tissue Segmentation results obtained using Model I (Thresholding Method) as described in PART A.
Fig. 8. Tissue Segmentation results obtained using Model II (BEP +Thresholding) as described in PART A.
rics (specifically TI) the best CSF, WM and GM tissues for each slice thickness are selected irrespective of the methods (as described in PART A) used for segmentation. Such 15 tissue samples of each type (CSF, WM and GM); that is, total 45 images are selected for training SVM. Testing is done with15 images (CSF-5, GM-5 and WM-5) of 5 subjects which are not used in training and are generated with Model I–IV. Experiments include- Linear SVM, RBF and Polynomial Kernel. 8. Results and discussions Figs. 7–10 shows the segmented Tissues (CSF, GM, and WM) for input MR scans of 1 mm slice thickness obtained with each of the Model I–IV respectively (as described in PART A). Table 1 shows the MI values for different BEPs, which is a measure of amount of information transferred from input images (combination of T1, T2, and PD) to band image. More the value of MI (in this experimentation, greater than 2), better will be the tissue information transferred to a band image. Such a band image will not
Method for Band Expansion
MI
DWT PCA Crosscorrelation Autocorrelation
1.79 2.17 2.19 1.79
Bold values indicate PCA and Crosscorrelation method has high values MI as compared to DWT and Autocorrelation.
be very similar to either of the input images (T1, T2 or PD), but will consist of unique combination of them based on the characteristic of the process used. With band images (in addition to original T1, T2 and PD scans) selected based on greater value of MI, the OC-ICA problem is overcome. At the same time, it is observed that such systematically selected input images for ICA, reduces the random generation of ICs (in every run, ICs generated will be in proper sequence). Table 2 shows the values of different quality indices that shows the comparison between each segmented tissue with the Ground Truth Image. As can be seen from Table 2, BEP +ICA is the only method for which TI has maximum value for all tissue types when compared with ground truth images. At the same time SI, Precision and Recall values for more number of segmented tissues are greater with BEP+ICA method as compared to segmented tissues obtained using remaining 3 methods. Thus, experimentation suggests BEP +ICA as best choice for brain MRI tissue segmentation into CSF, WM and GM. Most of the earlier work [5,6, and 8] suggests use of TI for comparison of segmented tissue with the ground truth. This experimentation (from Table 2), confirms the same and hence only TI is used for further comparison. For training SVM, the Tissue samples obtained using Model I–IV, with high TI are used. Table 3 shows the results for automatic tissue classification with SVM. (Training samples- 45; testing samples- 15). Here accuracy is calculated as the sum of correct classifications divided by the total number of classifications. SVM with Polynomial (order 3) gives accuracy of 100% for a complete set of test Tissue samples indicating the Polynomial kernel, which non-linearly maps input data into higher dimensional
46
R.S. Kamathe, K.R. Joshi / Biomedical Signal Processing and Control 40 (2018) 41–48
Table 2 Values of Different Quality Indices for 1 mm MRI scan: Segmented Tissue and Ground Truth Comparison. Quality Index »
Tanimoto Index
Method
CSF
GM
WM
Avg
Similarity Index CSF
GM
WM
Avg
Precision CSF
GM
WM
Avg
Recall CSF
GM
WM
Avg
Thresholding BEP ICA BEP+ICA
0.4 0.10 0.3 0.4
0.56 0.49 0.6 0.71
0.15 0.25 0.24 0.3
0.36 0.3 0.4 0.5
0.57 0.18 0.43 0.51
0.71 0.66 0.73 0.83
0.26 0.40 0.39 0.38
0.51 0.41 0.51 0.57
0.51 0.13 0.41 0.41
0.65 0.55 0.68 0.94
0.47 0.27 0.96 0.89
0.54 0.32 0.55 0.75
0.65 0.30 0.41 0.69
0.80 0.85 0.81 0.94
0.18 0.27 0.20 0.25
0.54 0.47 0.47 0.63
Bold values indicate highest values of different quality metrics for implemented methods, which reflects the best match between the classified tissue with its ground truth image.
Table 3 Automatic Tissue Classification with SVM Classifier. Test Images Obtained From different Models
Model I: Thresholding Model II: BEP Model III: ICA Model IV: BEP+ICA
Average Tissue Classification Accuracy (%) with SVM Linear
RBF
Polynomial
80 100 60 80
80 73.30 60 100
100 100 100 100
Table 4 Atrophy Ratio for AD Cases.
Fig. 11. Alzheimer’s disease Case: Image set- 1.
Fig. 12. Segmented Tissues obtained using BEP+ICA +Skull Removing + Thresholding Method (Model IV).
(feature) space, has improved separation margin for classes under consideration in feature space. 9. Case study: Alzheimer’s disease For case study of AD (Sample AD case is shown in Fig. 11), based on results obtained from ‘PART A’ experimentation, BEP (Based on cross correlation and PCA) + ICA (Model IV) is used for Tissue segmentation. Fig. 12 shows the segmented tissues for one of the AD cases (for AD case shown in Fig. 11). The Atrophy Ratio is calculated based on segmented tissues. Experimentation is carried out for 5 number of different AD patients [21]. Atrophy calculations for test image sets are as shown in Table 4. In this work, the axial scans of AD cases are used rather than Saggital scans used in [10]. Smaller the value of Atrophy Ratio more is the atrophy in brain indicating the progression towards AD. However, this study has
Image Set
Atrophy Ratio
1 2 3 4 5
0.7407 0.7309 0.6849 0.7809 0.7574
following limitations: i. very less number of AD cases is currently available for experimentation (As most of the databases for AD study comprises of only T1 scans e.g. ADNI [27] and OASIS [26].For ICA at least T1, T2 and PD scans are desired) and ii. Unavailability of previous test results on axial images. Due to this, the part of experimentation which aims to obtain most accurately segmented brain tissue samples (CSF, WM and GM) for AD cases, so that the atrophy ratio can be a correct representation of possible atrophy in AD cases, tissue segmentation results are validated based on Radiologist’s feedback. Table 5 summarizes the Radiologist’s feedback on Tissue classification provided by different Models (I–IV) used, on the scale of: Acceptable, Need Improvement, Not Acceptable- for the purpose of diagnosis. Tissue segmentation results for all four methods with dataset of 1 mm, 3 mm, 5 mm, 7 mm and 9 mm were shown to him during development stages. Table 5 comprises results for case study only. As can be seen from the table the BEP + ICA output is most acceptable for all set of images compared to other methods of segmentation. Thresholding though very simple gives acceptable segmentation of WM and hence can be used in diseases related to White Matter Lesions (WMLs). However, for predicting the threshold for Atrophy Ratio which can help to classify normal and abnormal (Atrophied brain cases), the experimentation with more number of AD and normal cases is required to be done. 10. Conclusion In this work we proposed the use of PCA and DWT based image fusion techniques for the use of band expansion for ICA and developed the automatic tissue classification system using SVM Classifier for brain MRI with scheme of using most accurate tissue classes validated with the help of different quality metrics and radiologist’s feedback. The polynomial kernel for SVM has given 100% accuracy in tissue classification as CSF, WM and GM for tissues extracted from
R.S. Kamathe, K.R. Joshi / Biomedical Signal Processing and Control 40 (2018) 41–48
47
Table 5 Radiologist’s Feedback Sheet. Methods
Acceptable CSF
Alzheimer Image Set 1 Thresholding BEP ICA BEP + ICA Alzheimer Image Set 2 Thresholding BEP ICA BEP + ICA Alzheimer Image Set 3 Thresholding BEP ICA BEP + ICA Alzheimer Image Set 4 Thresholding BEP ICA BEP+ICA Alzheimer Image Set 5 Thresholding BEP ICA BEP + ICA
√
√ √
√ √
√ √
√ √
Need Improvement GM
√
√
√
√
√
WM
Not Acceptable
CSF
GM
WM
√ √ √
√ √
√ √
√ √
√ √ √
√ √
√ √
√ √ √
√ √
√ √
√ √ √
√ √
√ √
√ √ √
√ √
CSF
GM
WM √
√ √
√ √
√ √
√ √
√
four different segmentation schemes, for test samples used. The BEP+ ICA has been applied successfully for case study of Alzheimer’s disease to detect possible brain atrophy. As the accuracy of BEP + ICA is most acceptable, upon having a set of images for Normal Persons and Persons with Mild Cognitive Impairment (MCI) (considered as early stage of Alzheimer’s or dementia); early diagnosis of Alzheimer’s disease will be possible. In future we would like to extend the work by overcoming the current problem of unavailability of all T1, T2 and PD scans of AD. Acknowledgments Authors are thankful to Dr. Rohit Sangolkar, DMRD, DNB (Radio), MBBS from Nizam’s institute of Medical Sciences, Hyderabad, India for verifying all the tissue segmentation results carefully and rating the results on the scale of ‘Acceptable’, ‘Needs Improvement’ and ‘Not Acceptable’. The segmented tissues using four different models implemented in study were shown to him. His valuable Inputs during the development stages for comparing outputs of implemented tissue segmentation methods, ‘manually’, with ground truth images has helped us to validate even the performance of quality metric selected for analysis. His valuable feedback in selecting most acceptable CSF, WM and GM tissue type along with statistical Quality metric value has helped us to train SVM with most accurate tissue samples leading to 100% accurate automatic tissue classification of test samples under consideration. References [1] A. Hyvärinen, Survey on independent component analysis, Neural Comput. Surv. 2 (1999) 94–128. [2] A. Hyvärinen, Fast and robust fixed point algorithms for ICA, IEEE Trans. Neural Netw. 10 (3) (1999) 626–634. [3] A. Hyvärinen, E. Oja, Independent component analysis: algorithms and applications, Neural Netw. 13 (4–5) (2000) 411–430. [4] A. Hyvärinen, Karhunen Oja, Independent Component Analysis, John Wiley and sons, 2001, pp. 1–12 (Chapter 1).
[5] Y.C. Ouyang, H.M. Chen, J.W. Chai, C. Chen, S.K. Poon, C.W. Yang, S.K. Lee, C.I. Chang, Band expansion based over-complete independent component analysis for multispectral processing of magnetic resonance images, IEEE Trans. Biomed. Eng. 55 (6) (2008) 1666–1677. [6] Y.C. Ouyang, H.M. Chen, C. Chen, S.K. Poon, C.W. Yang, S.K. Lee, Independent component analysis for magnetic resonance image analysis, EURASIP J. Adv. Signal Process. (2008), 780656, http://dx.doi.org/10.1155/2008/780656. [7] J.W. Chai, S.K. Lee, C. Chen, H.M. Chen, Y.C. Ouyang, SPIE Newsroom, 2009, http://dx.doi.org/10.1117/2.1200904.1498. [8] T. Nakai, S. Muraki, E. Bagarinao, Y. Miki, T. Takehara, K. Matsuo, C. Kato, H. Skahara, H. Isoda, Application of independent component analysis to magnetic resonance for enhancing of contrast of grey and white matter, Neuroimage 21 (2004) 251–260. [9] K. Vartak, R.S. Kamathe, K.R. Joshi, PCA, DWT based fusion as band expansion for blind source separation in magnetic resonance images, IEEE International Conference on Signal Processing, Communication Computing and Networking Technologies (2011) 581–586. [10] R. Sadek, Regional atrophy analysis of MRI for early detection of alzheimer’s disease, international journal of signal processing, Image Process. Pattern Recognit. 6 (1) (2013) 49–58. [11] C.A. Cocoso, A. Zijdenbos, N. Kabani, A. Evans, A Comparative Study of Model Based Brain MRI Tissue Classification Methods, BIC Technical Report, 2002, pp. 1–9. [12] D.W. Shattuck, R.M. Leahy, Brain suite: an automated cortical surface identification tool, Med. Image Anal. 6 (2002) 129–142. [13] C.A. Cocoso, A fully automatic and robust brain MRI tissue classification method, Med. Image Anal. 7 (4) (2003) 513–527. [14] S.M. Smith, M. Jenkinson, M.W. Woolrich, C.F. Beckmann, T.E. Behrens, H. Johansen-Berg, et al., Advances in functional and structural MR image analysis and implementation as FSL, Neuroimage 23 (2004) S208–19. [15] J. Ashburner, K.J. Friston, Unified segmentation, Neuroimage 26 (2005) 839–851, http://dx.doi.org/10.1016/j.neuroimage.2005.02.018. [16] H.M. Chen, J.W. Chai, S.K. Lee, C.C. Chen, Y.C. Lin, Y.C. Ouyang, C.I. Chang, W.C. Shen, Optimal parameters of SVM for classification of multispectral brain MRI, Proc. Int. Soc. Mag. Reson. Med. 17 (2009) 940. [17] J. Wang, C.I. Chang, Independent component analysis-based dimensionality reduction with applications in hyperspectral image analysis, IEEE Trans. Geosci Remote Sens. 44 (6) (2006) 1586–1600. [18] K. Kazemi, N. Noorizadeh, Quantitative comparison of SPM, FSL and BrainSuite for brain MR image segmentation, J. Biomed. Phys. Eng. 4 (1) (2014) 13–26. [19] Brain Web, Simulated Brain Database, and McGill University: http://www.bic. mni.mcgill.ca/brainweb/. [20] J.H. Morra, Z. Tu, L.G. Apostolova, A.E. Green, A.W. Toga, P.M. Thompson, Comparison of AdaBoost and support vector machines for detecting Alzheimer’s disease through automated hippocampal segmentation, IEEE Trans. Med. Imaging 29 (1) (2010) 30–43.
48
R.S. Kamathe, K.R. Joshi / Biomedical Signal Processing and Control 40 (2018) 41–48
[21] Harvard Medical School, http://med.harvard.edu/AANLIB/. [22] G. Geoff, Support Vector Machines and Kernel Methods, 2004. [23] R.M. Haralick, K. Shanmugam, I. Dinstein, Textural features for image classification, IEEE Trans. Syst. Man Cybernet. 3 (6) (1973) 610–621. [24] C. Corinna, V. Vladimir, Support-vector networks, Mach. Learn. 20 (3) (1995) 273–297.
[25] G. Qu, D. Zang, P. Yan, Information measure for performance of image fusion, Electron. Lett. 38 (7) (2002) 313–315. [26] Oasis Brain Database. 2012. http://www.oasis-brains.org. [27] ADNI Database. http://www.loni.edu.com.
Contents lists available at ScienceDirect
Biomedical Signal Processing and Control journal homepage: www.elsevier.com/locate/bspc
Research paper
A novel method based on independent component analysis for brain MR image tissue classification into CSF, WM and GM for atrophy detection in Alzheimer’s disease Rupali S. Kamathe a,∗ , Kalyani R. Joshi b a b
Department of Electronics and Telecommunication, College of Engineering, Pune, Maharashtra, India Department of Electronics and Telecommunication, PES’s MCOE, Pune, Maharashtra, India
a r t i c l e
i n f o
Article history: Received 4 April 2017 Received in revised form 26 July 2017 Accepted 5 September 2017 Keywords: Independent component analysis Band expansion process Brain MRI Support vector machines Alzheimer’s disease
a b s t r a c t Brain Magnetic Resonance Image (MRI) plays a vital role in diagnosis of diseases like Brain Tumor, Alzheimer, Multiple Sclerosis, Schizophrenia and other White Matter Lesions. In most of the cases accurate segmentation of Brain MRI into tissue types like Cerebro-Spinal Fluid (CSF), White Matter (WM) and Grey Matter (GM) is of interest. The diagnostic accuracy of expert and non-expert Radiologists can be improved with accurate and automated tissue segmentation and classification system. Such system can also be used for trainees to understand the individual tissue distribution in MRI scans. In this paper, we propose a novel automated tissue segmentation and classification method based on Independent Component Analysis (ICA) with Band Expansion Process (BEP) and Support Vector Machine (SVM) classifier which with input as T1, T2 and Proton Density (PD) scans of patient, provides output as CSF, WM and GM indicating the possible atrophy in brain which can help in diagnosis of Alzheimer’s disease (AD). The objective of this work is to test the effectiveness of ICA with different input images generated using BEP for accurate brain tissue segmentation by validating results with different quality metrics. The novel method for generating input images for ICA has been implemented and segmented tissues are used for atrophy detection. The BEP + ICA + Thresholding + ‘SVM trained with Grey Level Co-occurrence Matrix (GLCM) based texture features’ is giving 100% tissue classification accuracy for test samples under consideration. © 2017 Elsevier Ltd. All rights reserved.
1. Introduction Structural MRI is one of the important neuroimaging modality with high-resolution imaging and high brain tissue contrast capabilities. It is particularly used to look for brain tumors, stokes, blood clots, or other abnormalities that might account for Multiple Sclerosis (MS) or Alzheimer’s. MRI scans are acquired by different pulse sequences specified by three MR tissue parameters: spin–lattice (T1), spin–spin (T2) relaxation times, and Proton Density (PD) [5]. These are labeled as T1 weighted, T2 weighted, and PD weighted (Fig. 1). The constituents of brain, such as Gray Matter (GM), Cerebrospinal Fluid (CSF),White Matter (WM), Glial Matter, Fat, Muscle/Skin, etc. show unique characteristics under a magnetic field. However, the major tissue types of brain are CSF, GM and
∗ Corresponding author. E-mail addresses: [email protected] (R.S. Kamathe), [email protected] (K.R. Joshi). http://dx.doi.org/10.1016/j.bspc.2017.09.005 1746-8094/© 2017 Elsevier Ltd. All rights reserved.
WM; which have been distributed in T1, T2 and PD. As a result, spatial as well as tissue characteristics based features can be extracted from these MRI scans [6]. Reader based classification methods for these tissues are non-reproducible, and are practically difficult for the large amounts of data. Thus, development of fully automatic and accurate brain tissue classification from MRI in case of various disease symptoms like Tumors, MS, AD and other White Matter Lesions (WML) is of great interest [13,17] and is a challenging task. AD, the most common type of dementia is a major cause of disability worldwide. It can be detected at an early stage with the help of MRI so as to avoid irreversible damage of the brain with proper treatment plan. MRI can depict ‘atrophy’ − a decrease or shrinkage in the size of different areas of the brain caused due to wasting away of brain tissues in response to a disease process [10] like AD. Fig. 2 shows the various stages of Alzheimer’s disease. As can be seen in Fig. 2, MRI shows increasing atrophy as the disease progresses from MCI to AD (with no atrophy for healthy/normal controls). Measuring atrophy based on tissue volume distribution is important to diagnose, to monitor the disease
42
R.S. Kamathe, K.R. Joshi / Biomedical Signal Processing and Control 40 (2018) 41–48
Fig. 1. MRI Axial Scans.
Fig. 2. Normal Control (NC), Mild Cognitive impairment (MCI) and AD (T1 Weighted Axial Brain MR Images).
progression and in turn helps in the planning of treatment. Most of the segmentation methods are sensitive to the tissue intensity overlaps caused by limitations of MR image acquisition process (partial volume effect). There are three software packages widely used in neuroimaging community for structural and functional brain imaging study: FMRIB Software Library (FSL) [14], Statistical Parametric Mapping (SPM) [15] and BrainSuite [12]. FSL provides very effective tools like BET, FAST, FIRST, FLIRT etc. for structural MRI scans like T1-w or T2-w. FAST needs brain extracted images as input, which can be obtained using BET. SPM needs MATLAB and is based on a mixture model clustering algorithm. Kazmi K et al. in [18] presented detailed performance evaluation of these 3 methods. Chris et al. in [13] described and validated a completely automatic, non-parametric brain tissue classification system, based on a probabilistic anatomical atlas (model) for different tissue labels. Also, in [11] comparative study of different model based tissue classification methods is presented. Among other methods, ICA is also very popular [5,6,8,16]; often as a pre-processing step of SVM based classifier [6,16] and uses different types of structural MRI scans such as T1-w, T2-w and PD as an input. An application of ICA to MR image (MRI) analysis was investigated by Nakai et al. in [8] for contrast enhancement of GM and WM, till then it was in wide use for functional MRI (fMRI) [4]. In paper [8], study revealed that the involvement of grey or white matter in brain tumor cases and demyelination in the case of MS were enhanced and visualized in IC images. Also, in the same study [8], the potential of ICA for further analysis of anatomical images by enhancing the contrast among normal tissue or between normal and pathologic tissues has been investigated. In paper [6], Ouyang et al. pointed out two problems which were not addressed by Nakai et al.:i. Implementing Over Complete ICA (OC ICA) − number of image pulse sequences used for acquisition is generally smaller than the number of brain substances of interest and ii. ICA algorithm is initialized with initial random projection vectors to generate Independent Components (ICs). Because of this ICs generated are random and hence image evaluation cannot be performed until all ICs are generated. In [6], in order to mitigate this issue, the OC-ICA is used in conjunction with a feature extractionbased classifiers like- SVM and Fisher’s linear discriminant analysis (FLDA) for tissue classification. The performance of the method/algorithm for Tissue Segmentation can be validated by comparing the segmented tissue with Ground Truth image. For AD, for evaluating the disease progression, individual tissue volumes need to be measured and atrophy
should be measured with some parameter/s. Sadek in [10] has suggested three measures − Atrophy Ratio (AT), Alzheimer disease factor (ADF) and Progressive AD rate; for this purpose. In this paper, we present the automatic brain tissue classification based on ICA and SVM with novel method for generating input images (using BEP) for ICA to overcome limitations of ICA (mentioned in Section 2). The set of input images is selected with combination of original images (T1, T2 and PD) and band generated images (using different BEP); to generate independent Components (ICs). The approach is towards generating ICs which are as independent as possible with separate IC for each of the tissue of interest such as WM, GM and CSF. The idea for achieving accurate tissue classification is based on training of SVM with texture based spatial features evaluated on the most accurately segmented tissue samples (of WM, GM and CSF). Four different segmentation methods (Model I–IV) are proposed. The performance of all is tested and compared with different quality metrics like Mutual Information (MI), Tanimoto Index (TI) [16], Similarity Index (SI), Precision and Recall [20]. The paper also describes a case study for Alzheimer’s disease. In this study, the segmented tissues for AD cases are used further for calculation of Atrophy Ratio (AT). This paper is organized as follows: Section 2 introduces the ICA and its limitations. Section 3 is about BEP to create more inputs for ICA. Section 4 is about SVM classifier and GLCM Features. Section 5 describes the database used for experimentation. Section 6 gives the detailed methodology implemented in this work. Section 7 describes the various quality parameters used for quantitative analysis of methods implemented and results obtained. Section 8 is for results and discussions. Section 9 is about case study of AD followed by Conclusion in section 10. 2. Independent component analysis (ICA) ICA [1–4] is a linear transformation method. If ‘x’ denotes an m-dimensional random variable; the problem is then to find a function ‘f’ so that the n-dimensional transform s = (s1; s2; . . .. . .. sn) T defined by s = f (x) has some desirable properties and can be represented s = Wx where W is a matrix to be determined. The problems associated with ICA, of ICA being an overdetermined system with the number of samples (M) is usually greater than the sources to be separated (S) is pointed out in [6]. There were generally no solutions for such case. On the other hand, for M < S, ICA becomes an under-determined system. Nakai et al. [8] assumed that M ≥ S, where M in an MR imaging system consists of T1, T2, PD images [6,8] and their unique combinations. In [6] S is considered as the number of brain tissue substances, which includes water, blood, fat, GM, WM, CSF, and muscle, as signal sources to be separated; thus M < S and the ICA must deal with an over-complete representation of a mixed model, in which case many solutions are possible. In this work, Fast ICA: efficient and popular algorithm for ICA invented by Aapo Hyvärinen at Helsinki University of Technology [3,4] has been used. The classical application of the ICA model is that of blind source separation [1,7]; that is, no prior information about distribution of different tissues of interest in input images is required to be known. 3. Band expansion process (BEP) Ouyang et al. in [6] resolved the issue of having less number of input (band) images for ICA in MRI analysis by introducing the Band Expansion Process (BEP). The idea of BEP is to capture the correlation between original MR band images (T1, T2 and PD) and generate a set of second-order statistical band images [5]. The second-order statistics that can be used for BEP includes autocorrelation, cross-
R.S. Kamathe, K.R. Joshi / Biomedical Signal Processing and Control 40 (2018) 41–48
correlation, and nonlinear correlation [17] to create nonlinearly correlated images. In this work, we used fusion rules based on Principal Component Analysis (PCA) & Discrete Wavelet Transform (DWT) [9] for generating band images and found that the methods works well. Ouyang et al. pointed out in [6] that in most of the cases the band images obtained using cross correlation are generally sufficient to implement ICA. We also confirm the same; but the combination of band images obtained using cross correlation, PCA or DWT based methods with original T1, T2 and PD scans, as an input to ICA gives the best segmented tissues of interest. Our approach of generating band images using fusion based methods (as shown in Fig. 3) is inspired by the problem pointed out in [5] for auto correlated band images. Using auto correlated band images may sometimes cause nonsingularity problems in matrix computation because they are self-correlated and usually very close to the original images [5]. Fig. 3(a) and (b) respectively shows the PCA- based and DWT- based fusion scheme implemented for generation of band images. In Fig. 3(a), P1 and P2 are Principal Components of Images I1 and I2 respectively, which are combined by the rule P1I1 + P2I2. If I1is T1sequence, then I2 can be T2 or PD sequence. Likewise, different combinations of T1, T2 and PD sequences can be selected for both the methods. In DWT based fusion scheme, first order DWT is used. Then each of the sub-band pixels in LL, HL, LH and HH bands are combined using Pixel averaging to generate Fused Wavelet Coefficients. Inverse DWT is implemented to obtain the band image.
4. Support vector machine (SVM) and GLCM features SVM [22,24]: It often happens that the sets to discriminate are not linearly separable in finite dimensional space and hence there is need to map this space into a much higher-dimensional space, assuming that the separation will be easier in that space. A SVM [24] constructs a hyper plane or set of hyper planes in a high dimensional space, which can be used for classification. A good separation is achieved by the hyper plane that has the largest distance to the nearest training data points of any class (called functional margin). The SVM computes the simple dot product of the variables in the original space and may employ kernel tricks such as Gaussian Radial Basis Function, polynomial (homogeneous/inhomogeneous), Hyperbolic tangent etc. In this paper, multi-class SVM is implemented with 3 class labels as CSF, WM and GM. GLCM features: For feature extraction, method proposed by Haralick [23], namely, the Spatial Gray-Level Dependence method (SGLDM) (The corresponding matrix is labeled as GLCM- Gray Level Co-occurrence Matrix) is used. For each of the CSF, WM and GM image with N number of grey levels, the GLCM is a square matrix G of order N, where the (i, j)th entry of G represents the number of occasions a pixel with intensity i is adjacent to a pixel with intensity j. The normalized cooccurrence matrix is obtained by dividing each element of G by the total number of co-occurrence pairs in G. The adjacency can be defined to take place in each of the four directions (horizontal−0◦ , vertical−90◦ , left−135◦ and right diagonal−45◦ ). The Haralick’s texture features (Contrast, Correlation, Energy, and Homogeneity) are calculated for each of these directions of adjacency. These features are selected such that they show maximum similarity within the class and minimum similarity between the classes; so as to have good separation between classes under consideration (CSF, WM and GM).
43
5. Data set used for experimentation MRI Data set for Normal cases: McGill University [19]: Simulated Images, 0% noise and 0% INU, Stereotaxic Co-Ordinate System-(0.70,18) and Image slice Thickness–1 mm,3 mm,5 mm,7 mm and 9mm- 3 scans (T1, T2 and PD) of each case. Number of cases is 5. Thus 15 number of images available for a particular slice thickness (say 1 mm) comprising T1, T2 and PD for single experiment. Axial scans are selected as shown in Fig. 2. MRI Data set for AD cases: Harvard Medical School [21]: Real Patients’ data, Number of AD cases: 5; Axial scans are selected. No further pre-processing is required on selected T1, T2 and PD images. Training data for SVM: Four Haralick’s features as mentioned in Section 4 are extracted on each of the tissues classes- CSF, WM and GM obtained by segmentation of 15 number of input images (as described in section 5 (i)). Thus 60 numbers of training features are available for each of the tissue types- CSF, WM and GM respectively. 6. Quantitative analysis Mutual Information (MI) Metric [25]: used for BEP It gives the amount of information transferred from input image to Band Image. It is based on the measure of the degree of dependence of the two random variables A and B (any of the two input images T1, T2 and PD). Considering two input images A, B and F a new fused/band image, the amount of information that F contains about A and B can be calculated using MI. Following measures are used to compare, A: Extracted/segmented tissue and B: Ground truth of that tissue (shown in Fig. 5); to know how well the tissues of interest are segmented. Tanimoto index (TI) [16]: TI = A∩ B/A∪ B Given two sets A and B, if they are completely different, then TI = 0. TI=1 if the sets are completely the same (Tissue obtained after segmentation is exactly same as that of the ground truth). Similarity Index (SI) [20]: SI = |A∩ B|/| (A + B) /2| Precision [20]: Precision = |A∩ B|/|B| Precision gives how much amount of extracted tissue is precise to ground truth image. Recall: Recall = |A∩ B|/|A| Atrophy Ratio (AT) [10]: AT = (GM + WM) / (GM + WM + CSF) AT small values indicate highly atrophied brain. 7. Methodology PART A: Implementation of segmentation methods to obtain tissues of interest (CSF, WM, GM) Fig. 4 shows the four methods (named Model I–IV) implemented for segmentation of MRI scans (T1, T2 and PD) into CSF, WM and GM. For this experimentation, dataset as mentioned in Section 5 is used.
44
R.S. Kamathe, K.R. Joshi / Biomedical Signal Processing and Control 40 (2018) 41–48
Fig. 3. PCA and DWT Based Image Fusion as Band Expansion Process [9].
Fig. 4. Four Methods for Tissue segmentation.
Fig. 5. Automatic Tissue classification System for PART B.
Model I is based on simple Histogram based thresholding for segmentation of input MR Images into CSF, WM and GM. For Model II and Model III, input is not limited to T1, T2 and PD, but more number of input images is generated using band Expansion Processes (BEP), which includes images generated by performing − Autocorrelation, Cross correlation, PCA based Fusion and DWT based fusion on T1, T2 and PD MR Scans for experimentation. In Model II, thresholding is applied on these band images to generate CSF, WM, and GM. In Model III, 3 band images and 3 original images are given as an input to ICA to overcome the problem of OC-ICA. The ICs generated by Model III are further processed for skull removal and then,
thresholding is applied to obtain 3 tissues of interest- CSF, WM and GM. In Model IV, only 3 original MRIs- T1, T2 and PD are used as input to ICA to obtain 3 ICs. These ICs are processed further for skull removal and then, thresholding is applied to obtain3 ICs asCSF, WM and GM. PART B: Implementing Multi-Class SVM for automatic classification of segmented tissues into- CSF, WM and GM (Fig. 5) Tissue images (CSF, WM, GM) obtained using each of the above models (I–IV) described in PART A, are compared with Reference/Ground Truth images (shown in Fig. 6) with four parametersTI, SI, Precision and Recall. Table 2 shows the detail results for 1 mm slice thickness images. Same experimentation is repeated for 3 mm, 5 mm, 7 mm and 9 mm images. Based on the values of these met-
R.S. Kamathe, K.R. Joshi / Biomedical Signal Processing and Control 40 (2018) 41–48
45
Fig. 9. Tissue Segmentation results obtained using Model III (ICA+ Skull Removal +Thresholding) as described in PART A.
Fig. 6. Reference/Ground Truth Images [19].
Fig. 10. Tissue Segmentation results obtained using Model IV (BEP+ICA+Skull Removal +Thresholding) as described in PART A.
Table 1 MI Quality Metric.
Fig. 7. Tissue Segmentation results obtained using Model I (Thresholding Method) as described in PART A.
Fig. 8. Tissue Segmentation results obtained using Model II (BEP +Thresholding) as described in PART A.
rics (specifically TI) the best CSF, WM and GM tissues for each slice thickness are selected irrespective of the methods (as described in PART A) used for segmentation. Such 15 tissue samples of each type (CSF, WM and GM); that is, total 45 images are selected for training SVM. Testing is done with15 images (CSF-5, GM-5 and WM-5) of 5 subjects which are not used in training and are generated with Model I–IV. Experiments include- Linear SVM, RBF and Polynomial Kernel. 8. Results and discussions Figs. 7–10 shows the segmented Tissues (CSF, GM, and WM) for input MR scans of 1 mm slice thickness obtained with each of the Model I–IV respectively (as described in PART A). Table 1 shows the MI values for different BEPs, which is a measure of amount of information transferred from input images (combination of T1, T2, and PD) to band image. More the value of MI (in this experimentation, greater than 2), better will be the tissue information transferred to a band image. Such a band image will not
Method for Band Expansion
MI
DWT PCA Crosscorrelation Autocorrelation
1.79 2.17 2.19 1.79
Bold values indicate PCA and Crosscorrelation method has high values MI as compared to DWT and Autocorrelation.
be very similar to either of the input images (T1, T2 or PD), but will consist of unique combination of them based on the characteristic of the process used. With band images (in addition to original T1, T2 and PD scans) selected based on greater value of MI, the OC-ICA problem is overcome. At the same time, it is observed that such systematically selected input images for ICA, reduces the random generation of ICs (in every run, ICs generated will be in proper sequence). Table 2 shows the values of different quality indices that shows the comparison between each segmented tissue with the Ground Truth Image. As can be seen from Table 2, BEP +ICA is the only method for which TI has maximum value for all tissue types when compared with ground truth images. At the same time SI, Precision and Recall values for more number of segmented tissues are greater with BEP+ICA method as compared to segmented tissues obtained using remaining 3 methods. Thus, experimentation suggests BEP +ICA as best choice for brain MRI tissue segmentation into CSF, WM and GM. Most of the earlier work [5,6, and 8] suggests use of TI for comparison of segmented tissue with the ground truth. This experimentation (from Table 2), confirms the same and hence only TI is used for further comparison. For training SVM, the Tissue samples obtained using Model I–IV, with high TI are used. Table 3 shows the results for automatic tissue classification with SVM. (Training samples- 45; testing samples- 15). Here accuracy is calculated as the sum of correct classifications divided by the total number of classifications. SVM with Polynomial (order 3) gives accuracy of 100% for a complete set of test Tissue samples indicating the Polynomial kernel, which non-linearly maps input data into higher dimensional
46
R.S. Kamathe, K.R. Joshi / Biomedical Signal Processing and Control 40 (2018) 41–48
Table 2 Values of Different Quality Indices for 1 mm MRI scan: Segmented Tissue and Ground Truth Comparison. Quality Index »
Tanimoto Index
Method
CSF
GM
WM
Avg
Similarity Index CSF
GM
WM
Avg
Precision CSF
GM
WM
Avg
Recall CSF
GM
WM
Avg
Thresholding BEP ICA BEP+ICA
0.4 0.10 0.3 0.4
0.56 0.49 0.6 0.71
0.15 0.25 0.24 0.3
0.36 0.3 0.4 0.5
0.57 0.18 0.43 0.51
0.71 0.66 0.73 0.83
0.26 0.40 0.39 0.38
0.51 0.41 0.51 0.57
0.51 0.13 0.41 0.41
0.65 0.55 0.68 0.94
0.47 0.27 0.96 0.89
0.54 0.32 0.55 0.75
0.65 0.30 0.41 0.69
0.80 0.85 0.81 0.94
0.18 0.27 0.20 0.25
0.54 0.47 0.47 0.63
Bold values indicate highest values of different quality metrics for implemented methods, which reflects the best match between the classified tissue with its ground truth image.
Table 3 Automatic Tissue Classification with SVM Classifier. Test Images Obtained From different Models
Model I: Thresholding Model II: BEP Model III: ICA Model IV: BEP+ICA
Average Tissue Classification Accuracy (%) with SVM Linear
RBF
Polynomial
80 100 60 80
80 73.30 60 100
100 100 100 100
Table 4 Atrophy Ratio for AD Cases.
Fig. 11. Alzheimer’s disease Case: Image set- 1.
Fig. 12. Segmented Tissues obtained using BEP+ICA +Skull Removing + Thresholding Method (Model IV).
(feature) space, has improved separation margin for classes under consideration in feature space. 9. Case study: Alzheimer’s disease For case study of AD (Sample AD case is shown in Fig. 11), based on results obtained from ‘PART A’ experimentation, BEP (Based on cross correlation and PCA) + ICA (Model IV) is used for Tissue segmentation. Fig. 12 shows the segmented tissues for one of the AD cases (for AD case shown in Fig. 11). The Atrophy Ratio is calculated based on segmented tissues. Experimentation is carried out for 5 number of different AD patients [21]. Atrophy calculations for test image sets are as shown in Table 4. In this work, the axial scans of AD cases are used rather than Saggital scans used in [10]. Smaller the value of Atrophy Ratio more is the atrophy in brain indicating the progression towards AD. However, this study has
Image Set
Atrophy Ratio
1 2 3 4 5
0.7407 0.7309 0.6849 0.7809 0.7574
following limitations: i. very less number of AD cases is currently available for experimentation (As most of the databases for AD study comprises of only T1 scans e.g. ADNI [27] and OASIS [26].For ICA at least T1, T2 and PD scans are desired) and ii. Unavailability of previous test results on axial images. Due to this, the part of experimentation which aims to obtain most accurately segmented brain tissue samples (CSF, WM and GM) for AD cases, so that the atrophy ratio can be a correct representation of possible atrophy in AD cases, tissue segmentation results are validated based on Radiologist’s feedback. Table 5 summarizes the Radiologist’s feedback on Tissue classification provided by different Models (I–IV) used, on the scale of: Acceptable, Need Improvement, Not Acceptable- for the purpose of diagnosis. Tissue segmentation results for all four methods with dataset of 1 mm, 3 mm, 5 mm, 7 mm and 9 mm were shown to him during development stages. Table 5 comprises results for case study only. As can be seen from the table the BEP + ICA output is most acceptable for all set of images compared to other methods of segmentation. Thresholding though very simple gives acceptable segmentation of WM and hence can be used in diseases related to White Matter Lesions (WMLs). However, for predicting the threshold for Atrophy Ratio which can help to classify normal and abnormal (Atrophied brain cases), the experimentation with more number of AD and normal cases is required to be done. 10. Conclusion In this work we proposed the use of PCA and DWT based image fusion techniques for the use of band expansion for ICA and developed the automatic tissue classification system using SVM Classifier for brain MRI with scheme of using most accurate tissue classes validated with the help of different quality metrics and radiologist’s feedback. The polynomial kernel for SVM has given 100% accuracy in tissue classification as CSF, WM and GM for tissues extracted from
R.S. Kamathe, K.R. Joshi / Biomedical Signal Processing and Control 40 (2018) 41–48
47
Table 5 Radiologist’s Feedback Sheet. Methods
Acceptable CSF
Alzheimer Image Set 1 Thresholding BEP ICA BEP + ICA Alzheimer Image Set 2 Thresholding BEP ICA BEP + ICA Alzheimer Image Set 3 Thresholding BEP ICA BEP + ICA Alzheimer Image Set 4 Thresholding BEP ICA BEP+ICA Alzheimer Image Set 5 Thresholding BEP ICA BEP + ICA
√
√ √
√ √
√ √
√ √
Need Improvement GM
√
√
√
√
√
WM
Not Acceptable
CSF
GM
WM
√ √ √
√ √
√ √
√ √
√ √ √
√ √
√ √
√ √ √
√ √
√ √
√ √ √
√ √
√ √
√ √ √
√ √
CSF
GM
WM √
√ √
√ √
√ √
√ √
√
four different segmentation schemes, for test samples used. The BEP+ ICA has been applied successfully for case study of Alzheimer’s disease to detect possible brain atrophy. As the accuracy of BEP + ICA is most acceptable, upon having a set of images for Normal Persons and Persons with Mild Cognitive Impairment (MCI) (considered as early stage of Alzheimer’s or dementia); early diagnosis of Alzheimer’s disease will be possible. In future we would like to extend the work by overcoming the current problem of unavailability of all T1, T2 and PD scans of AD. Acknowledgments Authors are thankful to Dr. Rohit Sangolkar, DMRD, DNB (Radio), MBBS from Nizam’s institute of Medical Sciences, Hyderabad, India for verifying all the tissue segmentation results carefully and rating the results on the scale of ‘Acceptable’, ‘Needs Improvement’ and ‘Not Acceptable’. The segmented tissues using four different models implemented in study were shown to him. His valuable Inputs during the development stages for comparing outputs of implemented tissue segmentation methods, ‘manually’, with ground truth images has helped us to validate even the performance of quality metric selected for analysis. His valuable feedback in selecting most acceptable CSF, WM and GM tissue type along with statistical Quality metric value has helped us to train SVM with most accurate tissue samples leading to 100% accurate automatic tissue classification of test samples under consideration. References [1] A. Hyvärinen, Survey on independent component analysis, Neural Comput. Surv. 2 (1999) 94–128. [2] A. Hyvärinen, Fast and robust fixed point algorithms for ICA, IEEE Trans. Neural Netw. 10 (3) (1999) 626–634. [3] A. Hyvärinen, E. Oja, Independent component analysis: algorithms and applications, Neural Netw. 13 (4–5) (2000) 411–430. [4] A. Hyvärinen, Karhunen Oja, Independent Component Analysis, John Wiley and sons, 2001, pp. 1–12 (Chapter 1).
[5] Y.C. Ouyang, H.M. Chen, J.W. Chai, C. Chen, S.K. Poon, C.W. Yang, S.K. Lee, C.I. Chang, Band expansion based over-complete independent component analysis for multispectral processing of magnetic resonance images, IEEE Trans. Biomed. Eng. 55 (6) (2008) 1666–1677. [6] Y.C. Ouyang, H.M. Chen, C. Chen, S.K. Poon, C.W. Yang, S.K. Lee, Independent component analysis for magnetic resonance image analysis, EURASIP J. Adv. Signal Process. (2008), 780656, http://dx.doi.org/10.1155/2008/780656. [7] J.W. Chai, S.K. Lee, C. Chen, H.M. Chen, Y.C. Ouyang, SPIE Newsroom, 2009, http://dx.doi.org/10.1117/2.1200904.1498. [8] T. Nakai, S. Muraki, E. Bagarinao, Y. Miki, T. Takehara, K. Matsuo, C. Kato, H. Skahara, H. Isoda, Application of independent component analysis to magnetic resonance for enhancing of contrast of grey and white matter, Neuroimage 21 (2004) 251–260. [9] K. Vartak, R.S. Kamathe, K.R. Joshi, PCA, DWT based fusion as band expansion for blind source separation in magnetic resonance images, IEEE International Conference on Signal Processing, Communication Computing and Networking Technologies (2011) 581–586. [10] R. Sadek, Regional atrophy analysis of MRI for early detection of alzheimer’s disease, international journal of signal processing, Image Process. Pattern Recognit. 6 (1) (2013) 49–58. [11] C.A. Cocoso, A. Zijdenbos, N. Kabani, A. Evans, A Comparative Study of Model Based Brain MRI Tissue Classification Methods, BIC Technical Report, 2002, pp. 1–9. [12] D.W. Shattuck, R.M. Leahy, Brain suite: an automated cortical surface identification tool, Med. Image Anal. 6 (2002) 129–142. [13] C.A. Cocoso, A fully automatic and robust brain MRI tissue classification method, Med. Image Anal. 7 (4) (2003) 513–527. [14] S.M. Smith, M. Jenkinson, M.W. Woolrich, C.F. Beckmann, T.E. Behrens, H. Johansen-Berg, et al., Advances in functional and structural MR image analysis and implementation as FSL, Neuroimage 23 (2004) S208–19. [15] J. Ashburner, K.J. Friston, Unified segmentation, Neuroimage 26 (2005) 839–851, http://dx.doi.org/10.1016/j.neuroimage.2005.02.018. [16] H.M. Chen, J.W. Chai, S.K. Lee, C.C. Chen, Y.C. Lin, Y.C. Ouyang, C.I. Chang, W.C. Shen, Optimal parameters of SVM for classification of multispectral brain MRI, Proc. Int. Soc. Mag. Reson. Med. 17 (2009) 940. [17] J. Wang, C.I. Chang, Independent component analysis-based dimensionality reduction with applications in hyperspectral image analysis, IEEE Trans. Geosci Remote Sens. 44 (6) (2006) 1586–1600. [18] K. Kazemi, N. Noorizadeh, Quantitative comparison of SPM, FSL and BrainSuite for brain MR image segmentation, J. Biomed. Phys. Eng. 4 (1) (2014) 13–26. [19] Brain Web, Simulated Brain Database, and McGill University: http://www.bic. mni.mcgill.ca/brainweb/. [20] J.H. Morra, Z. Tu, L.G. Apostolova, A.E. Green, A.W. Toga, P.M. Thompson, Comparison of AdaBoost and support vector machines for detecting Alzheimer’s disease through automated hippocampal segmentation, IEEE Trans. Med. Imaging 29 (1) (2010) 30–43.
48
R.S. Kamathe, K.R. Joshi / Biomedical Signal Processing and Control 40 (2018) 41–48
[21] Harvard Medical School, http://med.harvard.edu/AANLIB/. [22] G. Geoff, Support Vector Machines and Kernel Methods, 2004. [23] R.M. Haralick, K. Shanmugam, I. Dinstein, Textural features for image classification, IEEE Trans. Syst. Man Cybernet. 3 (6) (1973) 610–621. [24] C. Corinna, V. Vladimir, Support-vector networks, Mach. Learn. 20 (3) (1995) 273–297.
[25] G. Qu, D. Zang, P. Yan, Information measure for performance of image fusion, Electron. Lett. 38 (7) (2002) 313–315. [26] Oasis Brain Database. 2012. http://www.oasis-brains.org. [27] ADNI Database. http://www.loni.edu.com.