Classification of Alzheimer's Disease MRI Images with CNN Based Hybrid Method

Alzheimer's is a neurological illness that occurs when brain cells are destroyed. Alzheimer patients suffer from confusion, difficulties in adapting to one's environment, problems related to speech and language skills, low motivation and problems. It is important to diagnose Alzheimer's illness early [1].

Studies on Alzheimer illness have been conducted in the literature. In their study, Lee et al. used deep learning methods for the diagnosis of Alzheimer's disease. They stated that when they used a single data method separately, they obtained an accuracy rate of 75%, and the second method they used achieved an accuracy rate of 81% [2].

Goo et al. applied more than one method in their study. They obtained accuracy values of 87.62% with CNN architectures, 85.61% with 2D SIFT method, 86.31% with 2D Kaze method, 85.26% with 3D SIFT method, and 83.15% with 3D Kaze method [3].

Zhao et al. stated that they obtained 92% accuracy by using SVM method in their study using 15 healthy and 15 patient data [4].

Ortiz et al. stated that using the deep learning methods, they increased the accuracy rate up to 90% in the diagnosis of the disease [5].

Moradi et al. stated that they achieved 90.2% accuracy using SVM and cross-validated. They used 10 folders for cross-validated. They stated that the results of their study played a major role in the diagnosis of the illness [6].

Salvatore et al. stated that they used a machine learning method that they optimized in their study for the diagnosis of Alzheimer disease and obtained the highest accuracy rate of 76% [7].

Lu et al. used the multiscale deep neural network structure to diagnose Alzheimer's illness. They stated that they used a data set taken from 1051 subjects and obtained an accuracy value of 82% [8].

In this paper, the Resnet50 architecture is used as the base and a hybrid model is proposed.

In this paper, CNN architectures are used to diagnose Alzheimer's illness. Results were obtained with Resnet50, Densenet201, Vgg16 and Alexnet architecture. Later, the classification operating was executed with the developed Hybrid model. The highest performance rate was achieved in the developed hybrid method. Resnet50 model was used as a base in the developed hybrid model. Some layers of Resnet50 model have been removed. In addition, new layers added to the Resnet50 model [9].

The article includes Theoretical Background in chapter 2, Application and Result in chapter 3 and Conclusion in chapter 4.

In this paper, it is aimed to classify using 4 class Alzheimer MRI images. First, the networks were trained with training data, and then they were tested using test data. While 80% of the data is used for training, 20% of the data is used for testing. The application was carried out on a computer with i7 processor in Matlab environment [14].

3.1 Performance metrics

There are several preferred methods for scale the performance of the classification. These calculations are calculated using the confusion matrix. The most preferred metrics are Accuracy, Precision, Recall and F-Score accuracy.

Confusion Matrix: The confusion matrix is one of the major metrics used in the classification process in CNN architectures. Other performance measures are calculated using the Confusion matrix [15]. Confusion Matrix used in the study is shown in Table 1.

Table 1. Confusion matrix

True Negative (TN): What is really negative is that the data is estimated negatively [17].

False Positive (FP): The fact that it is actually negative is that data is estimated positively [18].

False Negative (FN): The fact that it is actually positive is that data is estimated negatively [19].

The metrics obtained from the confusion matrix are defined for class C1. The same values are calculated separately for other classes. a1 expression belongs to class C1 and is the number of correctly predicted values.

TP= a1

TN= Sum of all values except values in row and column C1

FP=a2+a3+a4

FN=a5+a9+a13

Accuracy (ACC): Calculation of the accuracy value is given in Eq. (1) [20].

$A C C=\frac{T P+T N}{T P+T N+F P+F N}$  (1)

Error Rate (ERR): Calculation of the error rate is given in Eq. (2) [21].

$E R R=\frac{F P+F N}{T P+T N+F P+F N}$  (2)

F-Score: The F-Score value is presented in Eq. (3) [22].

$F-Score=\frac{2 * { Precision } * { Recall }}{ { Precision }+ { Recall }}$  (3)

Precision $=\frac{T P}{T P+F P}$  (4)

$\operatorname{Recall}=\frac{T P}{T P+F N}$  (5)

$S P=\frac{T N}{T N+F P}$  (6)

3.2 Experimental results

In this paper, it is aimed to classify the data in 4 different stages of Alzheimer disease. CNN architectures and the Hybrid method we developed were used for education. Also the performance of the architectures has been tested with the metrics that we have presented the theory of in the title.

Abbreviations used in tables;

A=MildDemented

B=ModerateDemented

C=NonDemented

D=VeryMildDemented

The accuracy and loss graphs of the network trained with the hybrid architecture we developed are presented in Figure 3.

When the network was tested with test data after training, the confusion matrix in Table 2 was obtained.

When the Confusion Matrix was examined, the hybrid model developed 90 test data correctly while placing 10 data in the wrong class. Accuracy rate in total is 90%. It achieved the highest accuracy rate in Class A with X%, while the lowest accuracy in Class B with y%.

3.jpg

Figure 3. Accuracy and loss curves of the hybrid model

Table 2. Confusion matrix of the Hybrid model

While the developed Hybrid model achieved the highest accuracy rate in the MildDemented class, it achieved the lowest accuracy in the ModerateDemented class.

The accuracy and loss curves obtained using the Alexnet architecture are presented in Figure 4.

Table 3. Hybrid model performance criteria

4.jpg

Figure 4. Alexnet accuracy and loss curves

Table 4. Alexnet confusion matrix

The Alexnet architecture, which was developed when Confusion Matrix was examined, classified 86 test data correctly and placed 14 data in the wrong class. The total accuracy rate is 86%.

The accuracy and loss curves obtained using the Densenet201 architecture are presented in Figure 5.

5.jpg

Figure 5. Densenet201 accuracy and loss curves

The confusion matrix obtained after the network is trained and tested with test data is given in Table 5.

Table 5. Densenet201 confusion matrix

The accuracy and loss curves obtained using the Resnet50 architecture are presented in Figure 6.

Resnet50 architecture, which was developed when Confusion Matrix was examined, correctly classified 78 test data and placed 22 data in the wrong class. The total accuracy rate is 78%.

The accuracy and loss curves obtained using the Vgg16 architecture are presented in Figure 7.

The confusion matrix obtained after the network is trained and tested with test data is given in Table 7.

The Vgg16 architecture, which was developed when Confusion Matrix was examined, correctly classified 78 test data and placed 22 data in the wrong class. The total accuracy rate is 78%.

The results obtained with the developed hybrid model and other CNN architectures are given in Table 8.

6.jpg

Figure 6. Resnet50 accuracy and loss curves

The confusion matrix obtained after the network is trained and tested with test data is given in Table 6.

Table 6. Resnet50 confusion matrix

7.png

Figure 7. Vgg16 accuracy and loss curves

Table 7. Vgg16 confusion matrix

Table 8. Accuracy rates of all models

Similar studies have been done in the literature before. But in most of the studies, 2 classes were used. It is either in 2 classes that he is sick or not. In this study, we have made an attempt to estimate the stage of the disease. The disease is classified in 4 stages in total. This is of very importance for the diagnosis of the illness. The previous studies related to the subject are presented in Table 9.

Table 9. Literature review table