Improving Efficiency in Prediction of Dementia Using Deep Learning Technique

Dementia [1, 2] is a disorder defined by a cluster of symptoms that impact daily functioning and can develop as a result of brain damage from disease or trauma. Memory, cognition, and behavior all gradually deteriorate as a result of the symptoms, which severely impact the sufferer's capacity to live a normal life. Dementia is defined by a gradual and severe decline in cognitive functions such as learning, memory, and thinking. Dementia, which can manifest in a variety of ways, is far from a natural aspect of becoming older; still, it impacts millions of people and becomes more prevalent with age (about one-third of those aged 85 and up may have dementia). Dementia progresses through seven identifiable stages, and the cognitive decline associated with dementia does not happen suddenly. The prevalence of Alzheimer's disease [3, 4] is high. Dementias like Alzheimer's begin with the dysfunction and eventual death of neurons, which are brain cells. Dementia patients experience an outsized loss of neurons compared to the general population as they age. Treatment may alleviate symptoms of certain diseases that mimic dementias, especially those caused by drug interactions or vitamin deficiencies.

It is now popular to use deep learning approaches that are based on feature selection, a type of data processing meant to supplement the built-in model selection procedure of conventional Deep Learning (DL) [5] methods. Every DL algorithm is known to have a method for choosing the best model from an ideal set of input features. The computational component of AI, uses data mining, pattern recognition, and knowledge discovery to automatically learn from data and create predicted task flows. Traditional artificial intelligence uses variables like (predictor and target variables) method. The training method attempts to discover a collection of model parameters that maximizes the association between the predictor (input characteristics) [6, 7] and the target variables by means of a looping procedure. The trained model employs the after receiving new input data, an established pattern can be used to estimate the target variable for the predictive factors. Patterns used in deep learning require at least two processors usually local computing settings, such as personal computers or cloud computing and corporate data centers. Prior to sending our data to an ML method, we make enormous efforts to alter them using simple linear correlation, or principal component analysis (PCA) analysis [8, 9]. Several experts automate the progression of the restrictions of the existing DL methods by using unusual approximations data to an DL method, we make enormous efforts to alter them using simple linear correlation, or principal component analysis (PCA) analysis. Several experts automate the progression of the restrictions of the existing DL methods by using unusual approximations. DL calculations are based on a difficult ML plan that is not exactly a match for the regular work method. We appear to have miscalculated the ML estimations since, for the most 8 parts, we ignore the subtleties of how objects go together and instead use them ready to move. HML is a development in machine learning that flawlessly combines several computations, methods, or procedures from comparable or dissimilar data spaces or use areas with the intention of improving one another.

No single DL [10] approach is suitable for all problems, just as no single cap fits all heads. Similar to how not one cap fits all heads, no singular ML approach is appropriate for all problems. When dealing with complex input spaces with multiple layers, some methods that perform well when dealing with noisy data may not be as effective. Others might not be the same as them. Due to the limitless possibilities for the hybridization of traditional DL approaches, everyone should be able to gather innovative combination models in a variety of ways. Various DL [11, 12] algorithms like CNN, RNN are used in recent times to predict various diseases to improve accuracy rather than ML algorithms.

An estimated 10 million new instances of dementia are reported annually, impacting over 50 million people [13] globally. The condition will be tripled by 20250 [14] which becomes very crucial problem among aged people. Using training and test data fed into a Convolutional Neural Network (CNN) Model, it is possible to identify people with mild dementia, moderate dementia, and severe dementia. The effort aims to diagnose Dementia early by utilizing MRI brain scan images for training with deep learning algorithms and then comparing the properties of demented brains to those of healthy brains to see if the former can more accurately predict illness.

The proposed system for detection of Dementia comes under the traditional image classification methods and the images are classified using the Deep Learning Technique [31], to be specific – Convolutional Neural Network is used to classify the images according to its classes.

Dementia detection approaches utilise convolutional neural networks, an essential part of Deep Learning. The model's procedure begins with collecting relevant data. After collecting a large quantity of MRI scan images, the following step is to divide the data into two sets: one for training and one for testing. It all starts with feeding the algorithm the pictures from the training dataset. The outcomes of the test set and training set are already known to us. The model is provided with the test data once it has retrieved features from the training data set. We can now see the metrics; accuracy and loss are the most crucial ones. Once the accuracy and loss are taken, we must train the model in such a way that it gives higher amount of accuracy, this can be done by changing the values in the hyper parameter where the model is trained. The model is now ready for validation, using the validation data set the model can be tested. The outcome is taken by feeding the validation data into the system.

The algorithm for the proposed model is given below in Figure 1.

The architecture diagram for GoogleNet CNN model is shown in Figure 2.

1.png

Figure 1. Algorithm for proposed model

2.png

Figure 2. Proposed CNN architecture

3.1 Data set

Biological MRI scans from both dementia and non-dementia patients are included in this dataset. The primary goal of the knowledge is to differentiate between healthy and Demented individuals mistreating the features, which is about zero for "healthy" and one for "Demented" people. The data The knowledge is in computational model of CNN. This dataset has the record of 5028 images of MRI scans, out of which 3200 patients have Dementia disorder, The dataset information is given below as Table 1.

Table 1. Dataset information

A number of characteristics were retrieved, similarly because the model will acquire the features of the MRI images by itself, the added advantage in using the convolutional neural network is the features need not be given to the system manually instead the training happens by reading and acquiring the characteristics in the image of binary class. In the training of the specified model, the dataset is differentiated into two classes, one is demented and the other one is healthy.

3.2 Data preprocessing

All the images in the training set are modified, but not the images in the test set, in order to avoid overfitting. The sole method to stop CNN from overfitting is to apply transformations to the images, which are nothing more complicated than simple geometrical transformations, zooms, or rotations. Therefore, to shift certain pixels, rotate the photos slightly, flip them horizontally, and zoom in and out, images are generally applied with some geometric changes. Many adjustments will actually be made to the pictures in order to enhance and modify them; this procedure is called image augmentation. .In reality, it involves adjusting the images in the training set to avoid overlearning in our CNN model. The images to the size target of (64, 64) is resized to (224,224), whether it was for the training set or test set since the GoogleNet CNN input size is set as (224,224) which will be compressed to 64 size output channel.

After Data preprocessing data exploration is done. Univariate graphical approach is applied to find out distribution's form and all of the data values are displayed in stem-and-leaf plots.

The object of the image is produced when each pixel takes a value between 0 and 255, which is particularly required for neural networks. The tool that does all the transformations to the training set's photos is represented by the Data Generator class named train datagen. For instance, the rescale parameter divides each pixel's value by 255 to apply feature scaling, while the subsequent changes amplify the images on the training set. Next, the train_datagen object is connected to the training set by importing it. Then traditional batch size 32 is selected to batch the photographs in each layer.

3.3 Building CNN

The following hyperparameters are set for CNN to predict dementia with highest prediction accuracy.

Three hidden layers are set up in one input layer, one output layer, and a 3 × 3 filter size is selected. The number of rows and the number of columns match the kernel size of the feature detector precisely. ReLU parameter is employed as activation function since it relates to rectifier activation. Max Pool is selected using a 3 × 3 pool size and stride 2 with 1 × 1 convolution layers in-between the max pool layers. The output of several pooling layers is flattened and pooled to produce a single-dimensional vector. An average pooling layer is created on top of that flattened layer instead of fully connected layer, and the input to the neural network will be a simple one-dimensional vector. CNN was trained with 25 epochs. With a factor of 10 applied per 5 epochs, the learning rate was set to 0.01.

Separation of the MRI images of brain into binary classes is a crucial step in the analysis. The output and the implementation of the overall process are shown with relevant datasets, processing details and suitable plots. The dataset containing the mixture of Demented and Non demented images are well trained into the Convolutional Neural Network, thus in the validation module our system is able to accurately judge whether the patient is suffering from Dementia or the patient is healthy. This convolution Neural network model is trained with 98% accuracy. The paper concludes the best accuracy was achieved on using GoogleNet CNN architecture rather than other CNN models. In future, the model shall be trained with even more accuracy to predict the results and the model to predict the accuracy with minimum optimal time.