Keywords

1 Introduction

Computer-assisted diagnosis and prognosis for brain disorders has always been of great interest in Alzheimer’s disease (AD) [1]. AD, a progressive, neurodegenerative disorder, is characterized by structural brain changes, leading to a gradual loss of cognitive functions [2]. Mild cognitive impairment (MCI) is an intermediate condition between normal ageing and dementia, distinguished by a cognitive decline greater than expected for a patient’s age and level of education, but which does not interfere with the patient’s daily life activities [3]. Some MCI patients do not convert to AD, and some even return to normal over time, but there is always a significant risk of AD conversion. In fact, more than half of MCI individuals do convert within 5 years [3]. MCI is therefore considered a substantial risk factor for AD.

Distinguishing MCI and cognitively normal (CN) controls from AD with intent to predict its onset or conversion has in recent years received a great amount of interest [4,5,6,7,8,9]. Multiple structural biomarkers, derived from magnetic resonance imaging (MRI), such as decreases in cortical thickness, in gray and white matter or subcortical volumes, have been extensively investigated for that purpose [10,11,12]. Multimodal studies involving various biomarkers, such as MRI or positron emission tomography (PET), resulted in high results for distinguishing between AD, MCI, and controls [10, 13]. Recent studies using deep learning approaches to brain imaging analysis have also obtained promising results within brain disorder diagnosis and prognosis [9, 13, 14].

The neuroanatomical abnormalities in AD have also been demonstrated to be reflected in the morphology of cortical sulci (Fig. 2a). They are considered as boundaries between various functional areas of the brain and are therefore related to its functional organization. Sulcal morphology is a promising neurological biomarker in AD and MCI [16,17,18,19]. Increase in sulcal widths [15,16,17] and reductions in depth [16, 19] have been observed in AD as a result of gray matter atrophy, when compared to normal ageing.

In this paper we postulated that a machine learning and pattern recognition approach involving a combination of sulcal and cortical features would result in higher AD classification results than if these features were used separately from each other. We hypothesized that these measures would be discriminative in classifying early MCI and AD, and that they would differ in their sensitivity in detecting the varying levels of brain atrophy in the elderly control subjects, in the early MCI subjects, and in advanced AD. Finally, we demonstrate the results of using either sulcal morphology or cortical thickness individually to distinguish between MCI, AD, and CN.

2 Methods

Sulcal and cortical features were extracted from T1-weighted MRI scans and used for classification of AD subjects from CN and MCI subjects, in a process illustrated on Fig. 1, using linear and Gaussian support vector machines (SVM). Additionally, we individually tested the classification performances of only sulcal features, as well as only cortical thickness features.

2.1 Data

Data used in the preparation of this article were obtained from the Alzheimer’s Disease Neuroimaging Initiative (ADNI) database (adni.loni.usc.edu). The ADNI was launched in 2003 as a public-private partnership, led by Principal Investigator Michael W. Weiner, MD. The primary goal of ADNI has been to test whether serial magnetic resonance imaging (MRI), positron emission tomography (PET), other biological markers, and clinical and neuropsychological assessment can be combined to measure the progression of mild cognitive impairment (MCI) and early Alzheimer’s disease (AD).

Fig. 1.
figure 1

Classification workflow: cortical sulci and cortical thickness measures were extracted from T1-weighted MR images to compute and select the most discriminating features for classification of AD and early MCI from CN.

Full size image

241 1.5T pre-processed T1-weighted magnetization-prepared rapid gradient echo (MP-RAGE) MRI scans of 85 CN, 75 MCI, and 81 AD subjects were acquired from the database. The demographics distribution for the three subject groups is presented in Table 1. The general inclusion/exclusion criteria were as follows: CN: Mini-Mental State Exam (MMSE) score between 24 and 30 (inclusive), Clinical Dementia Rating (CDR) of 0, non-depressed, non-demented, with no MCI. MCI: MMSE between 24 and 30 (inclusive), CDR of 0.5, general cognition and functional performance sufficiently preserved, negative AD diagnosis. AD: MMSE score between 20 and 26 (inclusive), CDR of 0.5 or 1.0; NINCDS/ADRDA criteria for probable AD.

Table 1. Demographic characteristics of CN, MCI, and AD subjects. The age, MMSE, and CDR scores are represented as mean and standard deviations.
Full size table

2.2 Sulcal and Cortical Feature Extraction

The T1-weighted MR images were first normalized and resampled to 1 mm\(^{3}\) voxels in SPM12 (Statistical Parametric Mapping software package for analysis of brain imaging data sequences). Then, cerebral sulci were extracted from the images using Morphologist 2013 pipeline included in the BrainVISA 4.4.0 software platform [20]. BrainVISA is a fully automatic anatomical segmentation pipeline, which produces a triangular mesh of the inner cortical surface of each brain hemisphere. Sulcal features were computed and extracted from 24 sulci in MATLAB R2018b, both the left and the right hemisphere from each of the 241 subjects (Fig. 3). Cortical thickness was computed in FreeSurfer (version 5.3.0). Thickness was computed as the average distance between the white matter and the pial surfaces (blue and green outlines, respectively Fig. 2b), along the normal vectors. In order to compute the sulcal morphology features, a medial surface was computed for every sulcus, in a process previously described in detail in our previous work [19]. In brief, sulcal meshes were computed as sets of three-dimensional vertices. Then we computed medial surfaces from sulcal meshes, consisting of a ridge and a fundus, as well as the set of new vertices located between the two faces of a sulcal mesh. For each sulcus, the following medial surface features were computed: length, mean depth, mean curvature, mean Gaussian curvature, and surface area. Fig. 4 illustrates the process of extracting the sulcal meshes, and computing the medial surfaces for subsequent feature extractions.

Fig. 2.
figure 2

(a) 3D view of the brain of a CN subject and the cortical sulci (blue). (b) the coronal view of the output volumes produced by FreeSurfer. The pial surface is shown as a green outline encapsulating the cortex, and the white matter is shown as the blue outline. (Color figure online)

Full size image

2.3 Feature Selection and Classification

In total, 310 features were computed for every subject: 70 cortical thickness features (34 from each hemisphere, one average thickness per hemisphere), and 240 sulcal features (mean depth, length, mean curvature, mean Gaussian curvature, medial surface area). Next, a feature selection process was applied in order to reduce their dimensionality, and select a feature combination that yielded the best classification results. Feature normalization was applied to standardize the features by rescaling them to the [0, 1] range (Eq. 1):

Fig. 3.
figure 3

The 24 cortical sulci extracted from each of the 241 subjects.

Full size image
$$\begin{aligned} X_{scaled} = \frac{X - \text {min}(X)}{\text {max}(X)-\text {min}(X)} \end{aligned}$$
(1)

where X was the original feature value, and \(X_{scaled}\) was the normalized one. Forward feature selection was performed, where each iteration of adding one feature to the SVM classifier was evaluated on the classification accuracy and a balance index B, defined in Eq. 2 (Garde et al. [21]), which permitted the extraction of features that maximized the classification accuracy, and at the same time provided a low difference between the true positives and negatives.

$$\begin{aligned} \textit{B} = \frac{\mid (1- \text {Specificity})-(1- \text {Sensitivity}) \mid }{[(1- \text {Specificity})+(1- \text {Sensitivity})]} \end{aligned}$$
(2)
Fig. 4.
figure 4

Representation of the steps included in the process of feature extraction (e.g. depth and length) from sulcal medial surfaces.

Full size image

SVM classifiers with a linear kernel and a Gaussian radial basis function kernel were used, with standardized predictor matrices, uniform prior probabilities for two classes, and a default value for the C parameter (C = 1). We selected a feature subset based on the SVM classifiers’ accuracy using 10-fold cross-validation from the set of 310 features. The best feature was added to the feature set, provided that this feature increased the accuracy, while simultaneously maintaining the balance index value under 0.4 [21]. All the other remaining features were individually tested in conjunction with the previously selected feature. This iterative procedure was repeated until the classification accuracy would no longer increase.

The classification procedure was divided into two parts: (1) feature selection identified a combination of sulcal and cortical features that distinguish between CN and AD in order to classify CN vs. MCI, and MCI vs. AD; (2) individual classifications using only sulcal and only cortical features. We performed ten realizations of 10-fold cross-validations. In each realization, the study population was randomly divided into ten separate folds, where each fold was used to test the classifiers’ performance, and the remaining folds were used as the training set. This procedure was independently repeated ten times, so that any bias possibly introduced by randomly partitioning the dataset in the cross-validation would be avoided.

3 Results

The cortical regions in which the atrophic changes were the most sensitive to classification between CN and AD are shown in Table 2. Six regions were identified in the left hemisphere, and two in the right. The left hemisphere has been reported in AD to be more severely affected by atrophy, and to lose gray matter faster than right, although a faster gray matter loss also occurs in age-matched healthy controls [2]. The selected features provided the highest CN vs. AD classification accuracy obtained with a linear SVM classifier (\(95.0\% \pm 0.92\) accuracy, \(93.0\% \pm 1.00\) sensitivity, and \(97.0\% \pm 1.30\) specificity, 0.95 AUC). The Gaussian kernel provided a higher AUC, but lower accuracy and specificity (Table 3).

Table 2. Cortical regions providing the highest distinction between the CN and AD groups using a linear SVM classifier. The CN, MCI, and AD values are represented as mean feature values and their standard deviations.
Full size table
Table 3. Classification results for CN, MCI, and AD subjects obtained with the two types of SVM kernels using ten realizations of 10-fold cross-validations. Results are shown as means and standard deviations.
Full size table

This feature combination was then applied to both the CN vs. MCI and the MCI vs. AD classifications. We achieved the highest average accuracy of 74.0% when classifying CN vs. MCI using a linear SVM kernel (67.0% using Gaussian kernel), and a similar result for the MCI vs. AD classification, also with a linear SVM (Table 3). The Receiver Operating Characteristic (ROC) curves for the three classifications using two SVM kernels are shown of Fig. 5: CN vs. AD on the left, CN vs. MCI in the middle, and MCI vs. AD on the right.

3.1 Classification with Either Sulcal or Cortical Features

Following the same feature selection procedure, we have identified two additional sets of features for CN vs. AD, CN vs. MCI, and MCI vs. AD classifications, in which only sulcal features and only cortical thickness features were used for classification. Table 4 shows the results using the linear and Gaussian radial basis function kernels, and the number of selected features. When used separately, cortical thickness features provided higher classification results than sulcal features for CN vs. AD, while sulcal features were more discriminating MCI from both CN and AD. However, the combination of both sulcal and cortical thickness yielded higher values of accuracy, sensitivity, specificity, and AUC for all three classification scenarios, than if these measures were used separately. The linear kernel SVM classifier provided higher classification results.

4 Discussion

In this paper we propose a machine learning and pattern recognition approach of integrating sulcal morphology and cortical thickness measurements for classification of AD and MCI. Linear and Gaussian SVM classifiers were trained with a selection of sulcal and cortical thickness features to classify CN vs. AD, CN vs. MCI, and MCI vs. AD. Additionally, separate sulcal and cortical features were individually tested for classification performance in these three classification scenarios.

Fig. 5.
figure 5

ROC curves for the CN vs. AD (left), CN vs. MCI (middle) and MCI vs. AD (right) classifications. Dashed lines represent random classification, blue curves represent the linear kernel SVM, and red curves illustrate the performance of the Gaussian kernel SVM. (Color figure online)

Full size image

SVMs with different kernels were chosen due to their high performances in high dimensional feature spaces [8, 22,23,24]. Cali et al. [23] reported no statistically significant differences between early-stage AD classification using Naïve Bayes, Logistic Regression and SVMs, but the SVMs outperformed the other classifiers with a combination of sulcal measures, cortical thickness, cortical volume, subcortical volumes and the MMSE score. Similarly, we obtained the highest classification results with linear SVM. Daliri [22] obtained a superior classification rate when using linear over Gaussian kernel SVM. Polynomial and sigmoid kernels provided even lower AD-classification rates, and thus were not implemented in our study. To evaluate the classifiers, we used 10-fold cross-validation, a statistical method for validating a predictive model by partitioning the original data into a training set, and a test set to evaluate its classification performance by averaging over ten iterations.

Table 4. Comparison between the classification results obtained with only sulcal features, only with cortical thickness measurements, and the combination of sulcal features and cortical thickness measurements.
Full size table

Our results confirmed that the combination of the sulcal and cortical measures were highly discriminative between CN and early MCI and AD subjects. The CN vs. AD classification yielded the 10-fold cross-validated, averaged accuracy of 95.0%, sensitivity of 93.0%, and specificity of 97.0% (0.95 AUC) with a linear SVM classifier. Suk et al. [9] used a deep ensemble learning of sparse regression models and obtained a maximum accuracy of 91.02% for CN vs. AD and 73% for CN vs. MCI classifications, using 10-fold cross-validation. Choi et al. [13] studied the brain metabolism using amyloid PET imaging and achieved an 86.6% accuracy with an SVM, and a 96.0% accuracy using a deep convolutional neural network, when classifying AD from controls. Beheshti et al. [25] used an approach for feature selection based on the t-test and a Fisher Criterion, using a voxel-based morphometry technique to compare the global and local gray matter differences in AD subjects and healthy controls using SVM. They reported a 96.32% accuracy using a 10-fold cross validation.

Furthermore, our results show that the features discriminating between CN and advanced AD provide relatively low accuracies when distinguishing early MCI from CN (77.2% accuracy with linear kernel), or advanced AD (73.3% accuracy with Gaussian kernel SVM). These inferior results may be due to the fact that the spread of atrophy in AD is hypothesized to be a non-linear process, and thus the structural cortical changes that distinguish AD from controls may be less sensitive to separate the controls from early MCI. Classification results reported in Table 3 suggest that the features characterizing an advanced atrophy of an AD patient are not as sensitive at detecting the differences between the brain of an early MCI subject and a control. They appear to be even less sensitive in distinguishing between early MCI and advanced AD.

Biomarker studies aiming at distinguishing between AD and normal controls often use either sulcal [15,16,17] or cortical features [11, 12]. However, the novelty of study is in combining these MRI-derived features, but also at analyzing them separately on the same data set, even though we acknowledge that some studies [13, 25] outperform our classification results. In our previous work [6] we aimed to identify the sulcal and cortical features to distinguish between stable MCI subjects and those converting to AD in order to identify features predicting AD-conversion. The separate feature selection process employed in this study revealed that sulcal features alone were superior to the cortical thickness at distinguishing early MCI from both AD and CN. This is likely due to the structural brain changes and regional volume losses being a natural part of normal ageing, occurring both in the cognitively normal elderly, and in AD patients. Therefore, cortical thickness alone may not be a sufficient biomarker, since a certain level of atrophy was already present in these two groups. In advanced AD, brain atrophy is highly pronounced, thus making cortical thickness a more sensitive biomarker (91.7% vs. 86.2%). However, sulcal features were the most sensitive for classifying early MCI from AD and CN. This indicates that sulcal morphology could be a potentially powerful biomarker in conjunction with cortical thickness for early-stage detection of neurodegenerative disorders.

5 Conclusion

The main contribution of this machine learning and pattern recognition study is a finding that a combination of sulcal morphology and cortical thickness measurements provides high classification results in discriminating between AD, MCI, and elderly control subjects. These results are competitive with the state-of-the-art techniques. Moreover, sulcal features were observed to be more sensitive at distinguishing MCI from AD and CN than cortical thickness, suggesting their potential as a structural biomarker for early detection of AD.