A Novel Approach of Diffusion Tensor Visualization Based Neuro Fuzzy Classification System for Early Detection of Alzheimer’s Disease

Research background and clinical significance

In our research work, we have undertaken the task of developing a screening test for the early detection of AD when cognitive features such as the intelligent behavior, power of rational thinking, and effective dealing with intellectual and higher integrative faculties are impaired. In this study, we have proposed a model of diffusion tensor visualization-based neuro-fuzzy classification system for early detection of AD [13]. We used the fuzzy (if–then) rules by taking three extracted features of ADC, FA, and GM values from the DMRI images of brain as input variables and evaluated the AD score as an output variable, which provides a useful platform to physicians in determining the status of the disease. The patterns were classified as either AD or healthy control patient. The experimental results show that out of the 20 samples, 14 for training, 3 for validation, and 3 for testing were used in the artificial neural network (ANN) classification system. The proposed network is trained and fed into the ANN with three input neurons, one hidden layer of thirty neurons, and two output neurons. The NN is trained with a training algorithm and the performance of the training algorithm is obtained by executing a fuzzy expert system (FES). The results show that the LM training algorithm gave a satisfactory performance for the diagnosis of AD, and it is an efficient tool as predictive accuracy equaled 100%.

DTI based neuro-fuzzy classification system of AD

A number of different tools or methods have been developed and investigated over the years as alternative screening tests for AD such as CT scan, traditional MR images [14], and DMRI of brain. They have served extensively for screening of AD patients or healthy controls. DTI-based neuro-fuzzy diagnostic systems of brain other than traditional tools have an important role in the cognitive assessment in brain white matter fibers tract integrity as a biomarker and in early detection of AD [15]. DTI provides information about anatomical microstructure changes in tissues with aging and aging-related cognitive decline in the brain of AD patients. DTI is an important tool to map the anisotropic water diffusion in the tissues. It visualizes the brain white matter tracts and tracks the diseased fibers or trauma. DTI can reveal abnormalities in white matter fiber structure or neurological disorder patterns in a variety of diseases including stroke, multiple sclerosis, dyslexia, schizophrenia, and tumor in brain white matter and provides a model of brain connectivity. FA is used to measure the rate of water diffusion in white matter tracts and characterize the DTI changes in each voxel of the tissue. It maps the axonal tracts and architecture and locates any injured part of the axon in myelin of the white matter. DTI provides the parameters such as FA and ADC to quantify the degree of preferred water diffusion restriction in different tissues. These findings may provide image methods to diagnose patients with early stage of AD.

Study population

We have used the clinical and laboratory database as ADC, FA, and GM of 20 patients from literature. The subjects utilized in this study were obtained from the Alzheimer’s Disease Neuroimaging Initiative (ADNI) database (http://adni.loni.usc.edu/). The study has been made for AD patients. One hundred and fifty brain scan images were downloaded from the ADNI database including the scans of AD patients (n = 95) and healthy controls (n = 55). Out of 150 patients (aged 65 to 96 years), we have taken 20 samples for the proposed work and experiment for classification of AD patients or healthy controls subjects (Tables 2 and 3). A summary of subject demographics and dementia status is shown in Table 1.

In this section, we reviewed this paper which has proposed a methodology for early detection of AD for 20 samples data set of patients from the literature using the measurement of ADC, FA, and GM of diffusion MR images of brain white matter [16]. These are used as input parameters into fuzzy tools, and using fuzzy rules, we evaluate the AD score in patients as an output variable which provides a useful platform to physicians in determining the status of the disease. In the second stage, the authors present an investigative study on AD and used the neuro-fuzzy classification system for pattern recognition of either AD or healthy control (Fig. 1).

Fig.1

Segmentation results of a normal individual and an AD patient. AD, Alzheimer’s disease; GM, gray matter; WM, white matter; CSF, cerebrospinal fluid.

From Table 2, the mean value and standard deviation of the ADC, FA, and GM values for AD patients (n = 95) are 1.97±0.82, 0.53±0.05, and 48.67±2.50, respectively; and for healthy control subjects (n = 55) are 0.69±0.25, 0.89±0.35, and 43.45±50.23, respectively. The actual values of ADC, FA, and GM for AD patients and healthy control subjects together is 150 each. Out of the 150–sample database, we have taken 20 samples database consisting of AD patients and healthy control subjects. The values of ADC, FA, and GM for individual patients were calculated from the mean value and standard deviation of ADC, FA, and GM from Table 2. It is shown in Table 3 where patient-III is the healthy control and patient-IX is the AD patient.

The proposed methodology for the AD classification process using the neuro-fuzzy system is shown in Fig. 2.

Fig.2

Proposed methodology for AD classification process by using neuro-fuzzy system. AD, Alzheimer’s disease; ANN, artificial neural network.

The procedures involving experiments on human subjects are done in accord with the ethical standards of the Committee on Human Experimentation of the institution in which the experiments were done or in accord with the Helsinki Declaration of 1975.

Preprocessing and segmentation of CT and MR images of AD patient

Preprocessing of images is required before segmentation of region of interest (ROI), feature extraction, and recognition. The processed images appear better in quality for diagnostics purpose and to distinguish more clearly the superimposed structures of the images. For certain specific application, the digital radiographs are modified by image enhancement and restoration to get more suitable results. Enhancement techniques convolve the gray level of a pixel independently of its neighbors using spatial masks. Enhancement includes image smoothing by two-dimensional filtering of images, sharpening, edge detection, and contrast stretching. Image segmentation is the process of localization of the ROI in the images. This is particularly true in radiographic image analysis where one needs accurate outlines of an object which is not too clearly defined in the original image. Edge segmentation of brain of AD patients plays an important role in recognition as edges of brain images holding much information of the brain image to get the information of the shape, size, and texture [17, 18].

Registration of CT and MR images of AD patient

The digitized images of axial sections of the same region of the brain of an AD patient obtained from T₁weighted MR, T₂weighted MR and CT modalities are used of pixel dimension 173 x 230. The three images under consideration are shown in Fig. 3a-c. The ventricular region (ROI) is shown in Fig. 4. The segmented edges of the ROI using canny edge segmentation [19] and their contours are shown in Fig. 5.

Fig.3

(a) T₁weighted MR image of brain of an AD patient, (b) T₂weighted MR image of brain of the same AD patient, and (c) CT image of brain of an AD patient.

Fig.4

Ventricular region of MR and CT images.

Fig.5

Contour of the ventricular region of MR and CT images.

Fuzzy expert system

Fuzziness is relevant to complexity because the nature of a system when its complexity exceeds a certain threshold, it becomes impractical or computationally infeasible to make precise assertions about it. Fuzzy control is based on fuzzy logic— a logical system which is much closer in spirit to human thinking and natural language than traditional logical systems. The fuzzy logic controller based on fuzzy logic provides a means of converting a linguistic control strategy based on expert knowledge into an automatic control strategy. The fuzzy inference system (FIS) has a vital role in the medical field to provide medical assistance to the radiologist to diagnose the abnormality in the medical images. Decision support systems through information technology has become a part of classification of AD [20].

In the first phase, we have used the process fuzzy (IF-THEN) rules by taking ADC, FA, and GM as input parameters and AD score as output variable. The second phase carries the classification of the AD score as either normal case or AD case by feeding the above input parameters to the FIS (Fig. 6).

Fig.6

Fuzzy inference system.

Input variables

2.5.1Apparent diffusion coefficient

This input variable has four fuzzy sets: “Very Small”, “Small”, “Medium”, and “High”. Membership functions include trapezoidal and triangular fuzzy numbers [21]. Fuzzy sets, support sets, and membership functions of ADC are identified in Table 4 and Fig. 7.

Table 4

Fuzzy set of input variable ‘Apparent diffusion coefficient’

Input field	Support set	Fuzzy set
Apparent diffusion coefficient	–1–1	Very Small
	0–6	Small
	3–9	Medium
	6–15	High

Fig.7

Membership function for apparent diffusion coefficient.

The membership functions of the fuzzy sets are given below:

μV.Small(x) = {1x≤0(1-x)    0 < x< 10    x≥1 μSmall(x) = {0   x≤0x   0 < x < 31   x = 3(6-x)/3   3 < x < 60   x ≥ 6μMedium(x) = {0    x ≤ 3(x-3)/3    3 < x < 61    x = 6(9-x)/3    6 < x < 90    x ≥ 9μHigh(x) = {0    x ≤ 6(x-6)/4.5    6 < x < 10.51    x = 10.5(15-x)/4.5    10.5 < x < 150    15 ≤ x

Fractional anisotropy

This input variable has four fuzzy sets: “Very Low”, “Low”, “Medium”, and “High”. Membership functions include trapezoidal and triangular fuzzy numbers. Fuzzy sets, support sets, and membership functions of FA are identified in Table 5 and Fig. 8.

Table 5

Fuzzy set of input variable ‘Fractional anisotropy’

Input field	Support set	Fuzzy set
Fractional anisotropy	0–0.5	Very Low
	0.25–0.75	Low
	0.5–1	Medium
	0.75–2	High

Fig.8

Membership function for fractional anisotropy.

The membership functions of the fuzzy sets are given below:

μV.Low(x) = {1    x ≤ 0.25(0.5 - x)/0.25    0.25 < x <0.50    x ≥ 0.5μLow(x) = {0    x ≤ 0.25(x-0.25)/0.25    0.25 < x <0.51    x = 0.5(0.75 - x)/0.25    0.5 < x < 0.750    x ≥ 0.75μMedium(x) = {0     x ≤ 0.5(x-0.5)/0.25    0.5 < x < 0.751    x = 0.75(1-x)/0.25    0.75 < x < 10    x ≥ 1μHigh(x) = {0    x ≤ 0.75(x-0.75)/0.25    0.75 < x < 11    1 ≤ x

Gray matter

This input variable has four fuzzy sets: “Very Small”, “Small”, “Medium”, and “High”. Membership functions include trapezoidal and triangular fuzzy numbers. Fuzzy sets, support sets, and membership functions of GM are identified in Table 6 and Fig. 9.

Table 6

Fuzzy set of input variable ‘Gray matter’

Input field	Support set	Fuzzy set
Gray matter	–100–20	Very Small
	–60–100	Small
	40–140	Medium
	90–200	High

Fig.9

Membership function for gray matter.

The membership functions of the fuzzy sets are given below:

μV.Small(x) = {1    x ≤ -60(20-x)/80    -60 < x 200    x ≥ 20μSmall(x) = {0    x ≤ -60(x + 60)/100    -60 < x < 401    x = 40(100-x)/60    40 < x < 1000    x ≥ 100

μMedium(x) = {0    x ≤ 40(x-40)/50    40 < x < 901    x = 90(140-x)/50    90 < x < 1400    x ≥ 140μHigh(x) = {0    x ≤ 90(x-90)/55    90 < x < 1451    145 ≤ x

Output variable

Alzheimer’s disease score

AD scores are calculated on the basis of the input variables ADC, FA, and GM fed into fuzzy inference system. This output variable has two fuzzy sets: “Low” (healthy control) and “High” (AD). Membership function include triangular fuzzy numbers. Fuzzy sets, support sets, and membership functions of AD scores are identified in Table 7 and Fig. 10.

Table 7

Fuzzy set of output variable ‘Alzheimer’s disease score’

Output field	Support set	Fuzzy set
Alzheimer’s disease score	0 –60	Low
	50 –100	High

Fig.10

Membership function for AD score.

The membership functions of the fuzzy sets are given below:

μLow(x) = {0    x < 0x/30    0 < x < 301    x = 30(60-x)/30    30 < x < 600    x ≥ 60μHigh(x) = {0    x < 50(x-50)/25    50 < x <751    x = 75(100-x)/25    75 < x < 1000    x ≥ 100

Fuzzy rule base

The fuzzy rule base consists of 16 rules that determine the AD scores (Low, High) by evaluation of the input variables ADC, FA, and GM, and it plays an important role in further treatment decision. The rule base of the system is shown in Table 8.

The linguistics variables such as Very Small, Small, Medium, and High of input variable ADC are determined. For the output variable AD score, the linguistics expressions are Low and High, respectively. For example, for Small (APC), Medium (FA), and High (GM), the membership functions will have to form respectively

The tested values of the system are shown in Table 9. Mamdani inference engine is used for rule evaluation purpose. The linguistic variables used in this system are like very small, small, medium, and high for input variables and low and high output variable. This study uses 16 rules in designing fuzzy rule base as shown in Fig. 11.

Fig.11

Developed fuzzy rules.

The rules have been developed using if-then method. Figure 12 show a rule view of this proposed design. For example, from Fig. 11, if ADC is high, FA is high, and GM is high, then output AD score is high. Using these rules, the result AD score in terms of percentage (%) has been computed. Surface views of this design are presented in Figs. 13–15. As also seen from Fig. 12, the value of AD score = 75. This means that the patients have AD score with a possibility 75%. Because each case is a quite high percentage and its tendency goes to severe AD stage, the doctor has to take necessary steps for the earliest possible treatment.

Fig.12

Calculation of the value of AD score for the values of apparent diffusion coefficient = –0.038, fractional anisotropy = 0.4076 and gray matter = 42.67.

Fig.13

Surface viewer of apparent diffusion coefficient and fractional anisotropy.

Fig.14

Surface viewer of apparent diffusion coefficient and gray matter.

Fig.15

Surface viewer of fractional anisotropy and gray matter.

Defuzzification

In this stage, truth degrees (α) of the rules are determined for each rule by aid of the min and then by taking max between working rules. For examples, for case-I where ADC = 0.966, FA = 0.4688, and GM = 45.37, Rule 2 and Rule 6 will be fired and we will obtain: α ₂ = min{Very Small (ADC), Low (FA), Small (GM)}  = min{0.034, 0.8752, 0.9105} = 0.034 α ₆ = min{Small (ADC), Low(FA), Small (GM)}  = min{0.966, 0.8752, 0.9105} = 0.8752. From Mamdani max-min inference, we will obtain the membership function of our system as max(α ₂,α ₆)  = max(0.034, 0.8752)  = 0.8752 that means Very High AD score. For case-II where ADC = 1.97, FA = 0.53, and GM = 48.67, Rule 6 and Rule 10 will be fired and we will obtain: α ₆ = min{Small (ADC), Low (FA), Small (GM)}  = min{1.97, 0.88, 0.8555} = 0.8555,α ₁₀  = min{Medium (ADC), Low (FA), Small (GM)}  = min{0, 0.88, 0.8555} = 0. From Mamdani max-min inference, we will obtain the membership function of our system as max(α ₆,α ₁₀)  = max(0.8555,0)  = 0.8555, that means Very High AD score. Then for both cases, we can calculate the crisp outputs. The crisp values of the AD score are calculated by the method center of gravity defuzzifier by the formula

Study population

For screening, we used here a 20 sample dataset of patients with ADC, FA, and GM as input variables from the literature, and risk status as output variables were evaluated using fuzzy rules to classify them as either AD or healthy control (shown in Table 9).

Neuro-fuzzy classifier and back-propagation learning algorithm

A neuro-fuzzy classifier is used to diagnose the AD patients at an early stage. The ANN is a network that interconnect all nodes. The input of a specific node is the weighted sum of the output of all the nodes to which it is connected. The output value of a node is a non-linear function, which is known as ‘transfer function’ or it is called ‘sigmoid function’ f(x) = 11+e-x of its input values. The nodes in the input layer receive the data and transfer them to nodes in the first hidden layer through the weighted links. The data are processed in the first hidden layer of the network and the result is transformed to the nodes in the next layer. The nodes in the last layer provide the network’s output. The j-^th node in a hidden layer processes the incoming data x _i by (i) calculating the weighted sum and adding a “bias” term θ _j as:

Design and architecture of neuro-fuzzy system

In this paper, we developed a FIS where input variables like ADC, FA, and GM are three clinical features obtained from patients fed into the network using fuzzy toolbox. We used NN toolbox from Matlab 2009 to evaluate the performance of the proposed networks where a three layer (one input, one hidden, and one output) feed forward network with three inputs and sigmoid hidden neurons and linear output neurons is suggested. The proposed NN is simulated with a set of 30 numbers of hidden neurons and the network is evaluated with a learning function LM in Fig. 16, which shows a model of the NN with 30 hidden neurons.

Fig.16

Neural network model.

Measures for performance

There are three performance measure, accuracy (AC), sensitivity (SE), and specificity (SP), which are used to evaluate the performance of the neuro-fuzzy classifier. These measures are calculated from the confusion matrix. SE and SP are the proportion of positive and negative cases that are well detected by the test. Classification AC depends on the number of samples correctly classified and they are defined as:

where TP is the number of true positive, FP is the number of false positive, TN is the number of true negative, and FN is the number of false negative. The confusion matrix is given in Table 9.

Experimental results

The experimental results show that the NN based on three extracted features of brain (ADC, FA, and GM) give a 100% classification rate. The confusion matrix of the clinical data is presented in Table 10. The confusion matrix of the predictive model of AD is shown in Table 11. The performance measures are calculated for the above clinical data of 20 patients from the NN that are shown in Figs. 17–19. The proposed network was trained and fed into the ANN with 3 input neurons, one hidden layer of 30 neurons, and 2 output neurons. Out of the 20 sample database, 14 for training, 3 for validation, and 3 for testing were used in the ANN classification system. In Table 9, the confusion matrix of the clinical data sets obtained the number of output of data sets of true positive (TP), false positive (FP), true negative (TN), and false negative (FN) as 9, 0, 11, and 0, respectively. The sensitivity, specificity, and accuracy are equal to 100% each. The NN is trained with a training algorithm and the performance of the training algorithm is obtained by executing an FES. The result shows that the LM training algorithm gives a satisfactory performance with 100% accuracy.

Fig.17

Results from neural network.

Fig.18

Confusion matrix of the neural network.

Fig.19

Receiver operating characteristic (ROC) curve.

CONCLUSIONS

In conclusion, the method of AD diagnosis is a successful model in assisting the physicians in the screening treatment of AD. In this paper, the presented FES successfully identified the presence of AD in the patients by measuring three parameters, ADC, FA, and GM, of the brain. These are considered as input variables, and the use of the neuro-fuzzy soft computing method can assist physicians in the detection of AD at early stages. From the analysis and diagnosis of the disease, the physician can decide which stage of AD the patient has and take the necessary steps for earliest possible treatment. This paper presents an investigational study for the classification of AD based of a neuro-fuzzy system for detection of healthy or AD patients. The experimental results have proven that the neuro-fuzzy based pattern recognition system gives a satisfactory performance. We intend to extend this methodology for early detection of Parkinson’s disease or other dementia.