Current Alzheimer Research, 2012, 9, 789-794
789
Computer based Classification of MR Scans in First Time Applicant Alzheimer Patients Fatma Polat#1, Selcuk Orhan Demirel#2,3, Omer Kitis3,4, Fatma Simsek3, Damla Isman Haznedaroglu3, Kerry Coburn5, Emre Kumral6 and Ali Saffet Gonul*,3,5 1
Beysehir State Hospital, Konya Turkey; 2Yasar University Department of Computer Engineering, Izmir Turkey; 3Ege University School of Medicine Department of Psychiatry SoCAT Lab Bornova Izmir Turkey; 4Ege University School of Medicine Department of Neuroradiology, Bornova Izmir Turkey; 5Mercer University School of Medicine Department of Psychiatry and Behavioral Sciences, Macon, GA, USA; 6Ege University School of Medicine Department of Neurology, Bornova Izmir Turkey Abstract: In this study, we aimed to classify MR images for recognizing Alzheimer Disease (AD) in a group of patients who were recently diagnosed by clinical history and neuropsychiatric exams by using non-biased machine-learning techniques. T1 weighted MRI scans of 31 patients with probable AD and 31 age- and gender-matched cognitively normal elderly were analyzed with voxel-based morphometry and classified by support vector machine (SVM), a machine learning technique. SVM could differentiate patients from controls with accuracy of 74 % (sensitivity: 70 % and specificity: 77 %) when the whole brain was included the analyses. The classification accuracy was increased to 79 % (sensitivity: 65 % and specificity: 93 %) when the analyses restricted to hippocampus. Our results showed that SVM is a promising tool for diagnosis of AD, but needed to be improved.
Keywords: Alzheimer’s disease, classification, diagnoses, support vector machines, hippocampus, magnetic resonance imaging. INTRODUCTION Alzheimer’s disease (AD) is the most common degenerative brain disease in the aging population. Although the disease and its pathological features in the brain were defined almost 100 years ago, diagnosis and treatment of AD are still clinical challenges. Even though, Diagnostic and Statistical Manual Criteria of American Psychiatric Association (DSMIV) and the National Institute of Neurological and Communicative Disorders and Stroke-Alzheimer’s Disease and Related Disorders Association (NINCDS-ADRDA) Work Group criteria have increased diagnostic accuracy, AD remains diagnosis of exclusion. Thus, a positive laboratory test capable of distinguishing between early stage AD and normal aging would be of considerable practical value. Structural brain imaging studies have described gray matter atrophy in AD patients compared to age-matched healthy controls. It was showed that histological changes and gray matter loss start before the clinically recognizable symptoms of AD and intensifies with memory problems as the disease progresses [1-4]. Depending on the stage of illness, the gray matter loss starts throughout medial temporal structures (hippocampus, entorhinal cortex, and amygdala) and the *Address correspondence to this author at the Ege University School of Medicine Department of Psychiatry SoCAT Lab, Bornova, 35100, Izmir, Turkey; Tel & Fax: +90 232 339 88 04; E-mail:
[email protected] The poster of this manuscript was presented at 23rd European Neuropsychopharmacology Congress in Amsterdam, 2010 and was awarded by Travel Award. #
These authors had equal contribution to the manuscript.
1875-5828/12 $58.00+.00
thalamus, later extending to parietal and frontal cortices. In comparisons (AD patients versus healthy controls), researchers have begun to focus their attention on distinguishing individual patients. The purpose of such studies is to assess whether various imaging methods are capable of categorizing patients and controls into their respective diagnostic groups with sufficient accuracy to be used prospectively in the diagnostic process. Studies using radiotracers like [(11)C]PIB have reported promising results. However, availability of these expensive tracers is limited, and radioactivity is an issue of concern to patients [5]. Structural magnetic resonance imaging (MRI) scanning is an ideal modality for assessment as a positive diagnostic test because of its wide availability, and an initial structural neuroimaging investigation is already recommended for exclusionary purposes as part of the differential diagnostic workup in AD [6]. Promising results separating patients from controls based on measuring specific brain areas like the hippocampus, entorhinal cortex and other medial temporal regions that are affected in the early stages of the disease have been reported [7, 8]. However, these region-specific measurements are time consuming and depend highly on the skill and anatomical knowledge of the user. On the other hand, multivariate analyses and supervised machine learning algorithms are becoming more popular for distinguishing patients from healthy controls because they can use whole 3-dimentional MRI image set and they are bias free. Support vector machine (SVM) software is chosen by many researchers as the supervised machine learning algorithm because it is developed from Statistical Learning Theory, and it shows considerable empirical performance in various fields like bioinfor© 2012 Bentham Science Publishers
790 Current Alzheimer Research, 2012, Vol. 9, No. 7
matics, text, and face recognition [9, 10]. Vemuri et al. [11] used linear SVM and reported 86% categorization accuracy for AD patients. Adding covariates like demographic variables and APO E4 genotype to the model increased the accuracy by another 3-4%. Also, using linear SVM, Kloppel et al. [12] correctly classified up to 95% of patients and controls. Davatzikos et al. [13] used pattern classification via nonlinear SVM and achieved 100% accuracy identifying probable AD patients and healthy subjects. In these studies, patients were recruited among the patients who were carefully followed up for a long time in Alzheimer Centers and/or the diagnoses were confirmed by post-mortem examinations. However, multivariate analyses and SVM has not been tested data for everyday clinic applicants. This group of patients is heterogeneous in the aspects of cognitive losses and stage of the disease. To test if the machine learning algorithms are successfully categories patients and cognitively normal people, we recruited a group of clinically diagnosed AD patients from first-time applicants to a neurology outpatient clinic and a group of age- and gender-matched healthy controls. MATERIAL AND METHODS Thirty-one patients (12 men, 19 women; mean age: 72.1 ± 7; MMSE: 19.8 ± 4.9; education: 6.68 ± 2.8 years) diagnosed as probable AD according to current NINCDSADRDA criteria were included in the study. Patients were either self-applicants for memory problems or were brought by their family members. All patients received a comprehensive diagnostic evaluation including medical histories, neurological and psychiatric examinations, and neuropsychological testing. The exclusion criteria for this study were a medical history of stroke, unstable cardiovascular disease or diabetes mellitus, seizure, alcohol abuse, head trauma, any neurological disease except AD, and psychiatric disorders. Patients with co-morbid diagnoses of mixed-type vascular dementia and AD were also excluded from the study. All patients were followed up for at least 6 months to increase diagnostic certainty. Thirty-one age- and gender-matched healthy controls with similar years of education (12 men, 19 women; mean age: 69.4 ± 6.8; MMSE: 28.5 ± 1.1; education: 7.6 ± 3.5 years) underwent the same diagnostic and screening procedures. We followed-up the same criteria for cognitively normal, healthy controls as Kloppel et al. [12]. Participants who had no symptoms at the initial clinical evaluation but during the workup those who had intracranial abnormalities, such as infarction, hematoma, or tumor (incidental meningioma) were excluded from the study. The imaging was performed on a 1.5 Tesla MR Unit (Magnetom Symphony; Vision to Symphony Upgrade, Siemens, Erlangen, Germany) with a circularly polarized head coil. The standard MRI scan included multiplanar turbo spinecho T1-weighted (T1-W) (TR/TE: 650/14ms) and T2weighted (T2-W) (TR/TE: 3800/90ms) images. In addition to conventional sequences, 3D FLASH sequence (TR/TE: 2300/3.93, slice thickness: 1 mm, voxel size: 1 mm3) in the sagittal plane was obtained for analyses. The DICOM (The Digital Imaging and Communications in Medicine) MR images were converted into NIFTI (Neuroimaging Informatics Technology Initiative) images and pre-
Polat et al.
processed with Statistical Parametric Mapping (SPM) version 8 software package (Wellcome Department of Cognitive Neurology, London, England). Images were segmented into grey matter (GM), white matter (WM), and cerebrospinal fluid by SPM. The DARTEL (A Fast Diffeomorphic Registration Algorithm) toolbox was used in order to register GM segments into population templates generated from all the images and then to normalize them into Montreal Neurological Institute (MNI) space. This non-linear warping technique minimizes structural variation between subjects. Details of this process can be found in [14]. A modulation step was used to ensure that the overall amount of each tissue class remained constant after normalization. No spatial smoothing was performed for SVM but an 8 mm kernel was used for group comparison by SPM. A hippocampus Region of Interest (ROI) mask was generated with Wake Forrest University (WFU)-Pick Atlas [15], an extension of SPM. Group comparison was done by General Linear Model where age, gender and total brain volume were the confounding factors. Conservative Family Wise Error (FWE) correction at =0.05 was used for multiple comparison. Principal Component Analysis (PCA) was applied to the segmented images by projecting the number of correlated voxels to a number of uncorrelated principal components (PCs). PCA reduces the complexity of classification caused by the high dimensionality of MR imaging data and the generalization error of the classification by optimizing the number of PCs for data projection, thus maximizing the degree of anatomical information while minimizing the impact of noise [16]. The PCA was performed by means of the Dimensionality Reduction Toolbox [17, 18]. The number of PCA components for this analyses was 54 and variance accounted for was 95%. The output of the PCA process was used to train a Support Vector Machine Classifier. Support vector machine (SVM) is a supervised multivariate classification method. SVMs are supervised in the sense that they learn about group differences in a training data set and apply the learned model to the classification of new data [19]. Technical information on SVMs can be found in [20]. Individual MR images are treated as points located in a high dimensional space. The total number of dimensions is determined by the numbers of voxels in each MR image. The use of an SVM for image classification is an example of a linear discrimination. In this model, SVM is a binary classifier which divides the space into which the MR images are distributed into two classes, AD and healthy, by identifying a separating hyperplane. In this work, LIBSVM software was used for SVM analysis [21]. The leave-one-out (“jackknife replication”) method was used to estimate the generalizability of the classification models. This method iteratively leaves successive images out of the training set for subsequent class assignment until each has been used in this way. In each iteration, the class membership of the test image is predicted by using the classifier constructed from the training data. RESULTS Compared to healthy controls, AD patients showed a wide array of atrophy mainly observed in the medial temporal areas including hippocampus and amygdala; precuneus,
Computer based Classification of MR Scans
Current Alzheimer Research, 2012, Vol. 9, No. 7
and whole cingulate cortex, parietal and frontal cortex (Table 1; Fig. 1). When we included the whole brain in the PCA, SVM could differentiate patients from controls with accuracy of 74 % (sensitivity: 70 % and specificity: 77 %) (Table 2). When we focused on hippocampus alone, the classification accuracy was increased to 79 % (sensitivity: 65 % and specificity: 93 %).
cortex. A highly similar pattern of gray matter loss for AD has been reported by previous neuroimaging studies that established group differences [22-25]. This pattern was also consistent with MRI findings of Braak Stage IV for neurofibrillary tangle distribution [26]. In recent years, there has been increasing interest in nonexpert-dependent, automated methods of clinical MRI classification for eventual development into a diagnostic aid in dementia. Although individual studies have used different methodologies, high-dimensional multivariate analyses followed by SVM classifications were the basic data processing strategies in several studies. Klöppel et al. [12] applied SVM to three different groups of AD patients and their counterparts. They achieved classification accuracies ranging from 81% to 92%. Vemuri et al. [11] reported 86% accuracy in a patient group with similar age and MMSE scores to our study group. Our accuracy of 74 % is smaller compared to
DISCUSSION In this study, we used automated classification of MRI images to distinguish individual AD patients from age- and gender-matched healthy subjects. As a group, our patients represent everyday geriatric patients who need medical evaluation for their dementia symptoms, and AD is one of the main differential diagnoses. When we compared the gray matter volume of AD patients with those of the healthy matched controls, we observed gray matter loss in medial temporal limbic areas, cingulate cortex, parietal and frontal Table 1.
Location and Talairach Coordinates of Gray Matter Loss in Alzheimer’s Patients
Talairach
Coordinates
P (FWE-cor)
K
Z
Location
x
y
z
10
-48
19
p
