Automatic Alzheimer's Disease Detection Based on Deep Learning Models

The death of healthy brain cells that leads to Alzheimer's disease (AD) is what causes a steady deterioration in mental, intellectual, and memory functions. It is the most typical cause of dementia, a condition that impairs social and mental abilities [1] This makes it difficult to go about your regular routine and gets worse with time. Nerve cell decline, neurofibrillary development and the gradual shrinkage of brain tissue all contribute to this occurrence [2]. The World Health Organization predicts that in 2021, there will be 55 million cases of dementia worldwide; by 2030, that number is expected to increase to 78 million, and by 2050, it will reach 139 million cases, more than twice as many as in 2021 [3]. While the prevalence of young-onset dementia is only 3% among young persons, which can be brought on by a variety of disorders, most individuals over 65 are at a high risk of developing dementia. Late treatment damages brain cells associated with cognition and memory, leading to loss of brain function, mental skill decline, language issues, and a reduction in the capacity to form coherent thoughts [4].

The illness begins with a slow degeneration of nerve cells and progresses to an acute stage of dementia that makes patients unable of doing basic everyday tasks [5].

As AD develops with time, it is crucial to diagnose it in its early stages. Two approaches are used to diagnose AD: (1) when the patient exhibits symptoms and (2) utilizing neuron imaging technologies. AD appears before any symptoms [6]. This phase, known as "pre-clinical AD," is characterized by the absence of symptoms in both the sick person and others around them Decades or more may pass during this phase of Alzheimer's disease. Whether or not symptoms are evident, amyloid beta deposits, a protein characteristic of AD, may be detected by contemporary imaging methods [7].

Because imaging methods provide a non-invasive interior inspection of the body, they are the most widely used method for identifying AD. When AD could only be identified after death, there was a challenging time in the history of AD therapy [8]. However, modern medical imaging technologies are increasingly crucial to the identification and management of AD. Numerous neuroimaging modalities, including computed tomography (CT), positron emission tomography (PET), functional magnetic resonance imaging (fMRI), and magnetic resonance imaging (MRI), are crucial in the identification of brain disorders [9].

Recently, machine learning is attracting the attention of academics in a variety of sectors due to its capacity of learn, enhance an algorithm and forecast the solution to any issue. Deep neural networks are able to use data to detect subtle and complex changes in brain structure, study how Alzheimer's disease progresses, and produce accurate results for the diagnosis of the illness [10]. CNN, auto encoders, recurrent neural networks and deep belief networks are only a handful of the numerous varieties of deep learning. While there are several deep learning models designed to work with two-dimensional image data, CNN is the most well-known of them. These models can also be used with one- and three-dimensional data. In order to train the model with optimal performance, a deep neural network requires a huge dataset [11].

In order to obtain an accurate diagnosis and save many lives, a methodology for detecting AD is suggested in this work. The suggested approach uses the MRI image diagnostic to categorize AD patients. High performance is the goal of this practice. In order to diagnose AD, we suggested a deep learning (DL) model based on the aforementioned concerns.

 The aim of this research is to present a convolutional neural network (CNN) based architecture that encompasses a range of processes, from image acquisition to AD classification. Using digital image processing, the proposed machine learning program classifies scanned MRI images and predicts the presence and severity of Alzheimer's disease.

As seen in section 2, many researchers have been developed various diagnostic strategies for AD. On the other hand, some research yielded good findings but required the implementation of a more involved and time-consuming model, while others did not. Therefore, the aim in this study is to minimize the amount of memory and processing power required for the diagnosis procedure while simultaneously improving the model and performance. The purposes of this paper are to: (i) diagnose AD as accurately as possible, (ii) update the previous model we proposed to improve accuracy, (iii) develop a new CNN model, (iv) comparison two suggested models in training and testing using different performance metrics; and (v) emphasize the effectiveness of different optimizers on the same model and compare them based on diagnosis performance. The suggested AD categorizing system show in Figure 1.

01.png

Figure 1. Block diagram of the proposed system

3.1 Dataset

This study's dataset was taken from Kaggle's Alzheimer's classification dataset. 1279 and 5121 scans, respectively, make up the training and testing portions of the dataset. This dataset information is show in Table 2.

Table 2. Information of Alzheimer’s dataset

52 Moderately Demented, 2560 Non Demented, 717 Mild Demented, and 1792 Very Mild Demented categories. The file format is jpg, and the image dimensions are 176 × 208. It is a valuable on the website: https://www.kaggle.com/datasets/tourist55/alzhe imers-dataset-4-class-of-images.

3.1.1 Dataset pre-processing

Preprocessing involves two stages for images: transformation and normalization. During the transformation process, the input image is resized to 224 × 224. Because the model learns more quickly on smaller images, resizing the image is an essential step in the model training process. Normalization is done to each of them following the scaling of the image. The method of normalization is used in this work is max-man method. We adjust the image to normal. The process of normalizing an image involves rescaling each pixel from a given range (0 to 255) to a value between 0 and 1. The updated images are saved in a pickle file once the transformation and normalization have been completed.

3.1.2 Dataset augmentation

The imbalance problem is addressed by the data augmentation process. To address overfitting and image inconsistencies, the proposed technique generates augmented images from original slices using six operations (rotation, translation, gamma correction, random noise addition, scaling, and random affine transformation). We apply augmentation techniques with the following configurations: shearing, zooming, horizontal flipping, 0.1 and 0.2 range width and height adjustments, rotation at 15 degrees. We improved and randomly duplicated photos between classes to balance the collection and increase classification accuracy. Following dataset expansion, Mild Demented has 2519 image, Non Demented 2561, Very Mild Demented 3584, and 364 Modified Demented. The initial training dataset was expanded with the addition of the supplemented data to provide a sufficiently enough sample size.

3.2 The first proposed CNN model

In the proposed CNN model, the images are resized to (224 × 224), with three dimensions. This section will address the construction of the best-proposed model, which consists of four convolutional layers with max-pooling and (0.2) dropout after each conv2D layer. Initially, entered 16 filters convolutional layers were used with a (3 × 3) filter and stride its (1 × 1) with the same padding. Then, a max-pooling layer with a 2 × 2 filter was added to get the image reduced in dimensions with (2 × 2) strides and add dropout with (0.2); at the same point, convolutional layers and filters were increased to 32, 64, and 128 with the same filter size of (3 × 3).

Finally, a dense layer of 1024 neurons and a softmax output layer were used to determine probability scores for each class. Figure 2 shows the basic layers of the proposed CNN model. The run time 7s.

02.png

Figure 2. Basic layers of the proposed CNN model

Table 3 illustrates the parameters for each layer of the proposed model. Before training, several hyperparameters may be specified, including the loss function (binary cross-entropy), Adam optimizer, epoch (200), learning rate (0.001), activation function (softmax), kernel regularizer (0.01) is used during model training for preventing overfitting, and batch size (32). The network contains 25,790,626 parameters that may be learnt.

Table 3. Parameters for each layer in the proposed CNN

3.3 The second pre-trained DenseNet169 model

While it may be used to a variety of datasets, the pre-trained model was developed using the Image Net dataset. Our strategies include fine-tuning and transfer learning. To avoid information loss during transfer learning, a model is trained on a large number of small datasets before freezing the layer. To retain feature extraction, we freeze the convolution and pooling layers and add a new classifier layer that fits our dataset. Using the dataset, we train the layer that recognizes complex patterns, and we freeze the layer that recognizes generic features. Therefore, we used a transfer learning strategy called fine-tuning. Freezing every convolution and pooling layer is not necessary for fine-tuning. The pooling and final convolution layers are no longer frozen.

Table 4 shows that just the last convolution and pooling layer was retrained, whereas three fully connected layers were added. The reasons of transfer learning strategy selection, it gives best result from other methods.

Table 4. Parameters of each layer for the pre-trained DenseNet169 model

The overview of the second suggested DenseNet169 model is displayed in Figure 3 and Table 2. This model was trained using a 224 × 224 preprocessed image. Our data are divided as follows: 20% is for testing, 20% is for validation, and 60% is for training. The dataset with 15 rotation range, 0.1 width and height shift, 0.2 shear, and 0.2 zoom factor is subject to the augmentation procedure. The number of epochs and bach size are used (200 and 32). Adam optimizer achieved the best effective on the data.

03.png

Figure 3. The architecture of the DenseNet169 model