Alzheimer’s disease diagnosis and classification using deep learning techniques

Introduction

Alzheimer’s disease (AD) is a type of incurable brain disease due to neurodegeneration. AD is present worldwide. AD is characterized by β-amyloid (Aβ), which contains extracellular plaques and tau-containing intracellular neurofibrillary tangles (Knopman et al., 2021). Cognitive ability disorder is the major symptom due to AD. This disease is more prevalent in aged people, normally affecting those aged 65 or older; 10% of cases are early onset occurring in people younger than 65. AD also affects language, attention, comprehension, reasoning, and memory. Professionals can take care of patients suffering from this disease’s symptoms. The decline in cognitive ability occurs due to dementia which impacts daily activities. AD is the most common type of dementia, accounting for about two-thirds of the cases due to age factors. In 2020, AD was the seventh leading cause of death in the United States of America. AD has some treatments to improve the symptoms, but there is no proper treatment to recover (Kumar et al., 2018).

AD types are classified as Non-Dementia, Very-Mild Dementia, Mild Dementia, and Moderate Dementia. According to AD analysis, these stages are defined according to other research techniques and are different from the DSM-5 classification of AD (Kumar et al., 2018). Symptoms of AD depend upon the stage of the disease. The most common and very first symptom is short-term memory loss. Also, a language disorder is common in AD patients. AD symptoms normally do not show in the early stages, which is the major barrier to proper treatment. As there is no proper treatment for AD, early diagnosis enables potential treatment to cover the early stages. However, early diagnosis is challenging due to the presentation of minor symptoms and sometimes it is not possible to properly identify the symptoms. Normally, a neuropsychological examination is used for the early diagnosis of AD. Clinicians are responsible for properly analyzing the AD patient, but the manual analysis is tedious and takes time for the number of patients symptomatic for the disease (Salehi, Baglat & Gupta, 2020; Butt et al., 2019).

Medical experts are responsible for AD diagnosis, ideally found in the early stages. However, unfortunately, due to the large volume of data in medical images, and the large number of patients, it is impossible to accurately and quickly analyze them manually. Every clinician or medical expert manually examines a few of the records and provides an analysis based on their experience and knowledge. The risk of improper analysis may cause more problems. There is a need for an automatic solution for the analysis of the large volume of imaging data from patients. Various techniques, such as the Internet of Medical Things (IoMT), clinical treatment in labs using MRI or CT images, machine learning-based systems for analyzing large volume data, and deep learning approaches are playing an important role in the medical domain (Kundaram & Pathak, 2021). MICCAI BRATS challenges also provide solutions using deep learning and machine learning approaches on MRI or CT images (Frozza, Lourenco & De Felice, 2018).

Background

AD is a fast-growing disease occurring worldwide. It mostly affects the aged population. AD is incurable and is a neurodegenerative disease that mostly affects the brain. People are facing problems associated with AD due to a lack of early diagnosis due to no or minor symptoms in the early stages of the disease. Gaudiuso et al. (2020) proposed a technique using machine learning integrated with Laser-Induced Breakdown Spectroscopy (LIBS). Micro drop plasmas were diagnosed for AD patients and healthy controls (HCs). The classification was also performed using machine learning algorithms. The dataset used for the evaluation of the model had 31 AD patients and 36 HCs. The proposed technique successfully diagnosed late-onset AD, with a better diagnosis for patients greater than 65 years of age. The total classification accuracy between AD and HC was 80%, which indicates that it is better than other approaches.

Bilal et al. (2020) proposed a technique for the accurate diagnosis of AD. This study proposed nanotechnologies to overcome the limitations in treatment for early diagnosis and analysis. The neurodegenerative disease was treated with nanocarriers delivering bio-actives. Using nanocarriers with bio-actives is a very fruitful way to treat neurodegenerative disease compared to common therapies. This review came after studying nanocarriers, nanoparticles, and nanotubes which are helpful for early-stage diagnosis in large volumes of data. This study conducted a few experiments to find an optimum solution for early diagnosis for aged people to overcome AD. The results of this study perform better than the open-source dataset and suggest how to treat patients in the early stages.

DeTure & Dickson (2019) proposed a method for the diagnosis of AD, specifically neuropathological diagnosis for AD. AD is the most common disease and a cause of dementia, affecting memory. This study reviews various ways to find an improved solution for the diagnosis of AD. The major problem identified is how to manage the clinical treatment of many patients, especially when there is no properly defined treatment for the disease. Artificial Intelligence and machine learning also play their role in monitoring a large number of patients in a short time, specifically for the early diagnosis of the disease. It is very difficult to diagnose the disease based on symptoms, which include various effects on memory and conversation, and there is a need to address the problems for early diagnosis. This study gives a solution for diagnosis in early AD stages using machine learning approaches.

Chitradevi & Prabha (2020), proposed a solution for diagnosing AD by analyzing brain sub-regions using deep learning approaches and image processing optimization techniques (Kaggle, 2019). The deep learning algorithms used were Genetic Algorithm, Partial Swarm Optimization, Grey Wolf Optimization, and Cuckoo Search algorithm. Grey Wolf Optimization outperformed and was more stable than the other algorithms. Deep learning algorithms were used as validation tools to classify normal and abnormal AD. Among all the sub-regions of the hippocampus is the important region where the maximum chance of detecting AD is found in the early stages. This study evaluated using an open-source dataset and obtained values for accuracy, sensitivity, and specificity of 95%, 95%, and 94%, respectively. This score shows the proposed study is most correlated with the Mini-Mental State Examination.

AD diagnosis has been performed using different deep learning and machine learning algorithms on the image dataset, whereas Sharma, Vijayvargiya & Kumar (2021) compared a few deep learning techniques, specifically convolutional neural network (CNN) architectures, for AD. This study analyzed eight transfer learning techniques for the four classes of AD: Non-Dementia, Very Mild-Dementia, Mild-Dementia, and Moderate Dementia. Eight transfer learning approaches were tested: DenseNet-169, MobileNet-V2, ResNet-101, Inception-V3, ResNet-50, VGG-16, VGG-19 and Inception ResNet-V2. The highest score for evaluating these techniques was valued for accuracy and precision at 98.0% and 98.1%, respectively, with these being the best scores in the medical domain. This study proves that using these techniques for AD diagnosis is effective. A deep feature-based real-time model was introduced (Khan et al., 2020) to predict the AD stage. This technique used CNN, k-nearest neighbours (KNN), and support vector machine (SVM) to classify the AD stage from the image dataset and obtained outstanding performance with 99.21% accuracy values. That proves the performance of this research study.

The state-of-the-art studies clearly show the needed system for the early diagnosis of Alzheimer’s disease. The major contributions of the proposed model are mentioned.

Materials and Methods

Proposed methodology

Figure 1 presents this study’s research methodology for finding a solution to AD’s accurate, early diagnosis. This proposed research methodology addresses the problems discussed in the Introduction. Various techniques based on deep learning were discussed earlier, but these approaches are lacking in the early diagnosis of AD when symptoms are minor or non-existent.

This research study focused on the CNN-based deep learning models known as DenseNet169 and ResNet50, which are used for diagnosing and classifying AD.

Model building

The proposed model was based on the DenseNet169 and ResNet50 CNN models. CNN is a deep learning model used for the classification of features. A CNN model contains several different layers. A few common layers in CNN architectures are discussed below.

The input layer

The first CNN layer is the input layer which defines the image size used in the dataset as 240 * 240 * 3 (width * height * channels). Shuffling the images is unnecessary because each training epoch will be shuffled automatically.

The convolutional layer

The convolutional layer is the core of the CNN architecture. It is the basic master layer which contains required parameters and feature maps. These layers are key to performance via the selection of a kernel. Padding is used in the feature maps and the convolutional layer to match the sizes of the input and output layers. By default, padding is assigned pre-set values of one.

Batch normalization layer

Network training is a time-consuming process. Batch normalization of the training data is an easy optimization. Gradients are normalized and help move forward or trigger the network propagation. Batch normalization layers are used between the non-linearity and convolutional layers.

ReLU layer

Rectified linear unit (ReLU) is the activation function mostly used in neural network architectures. ReLU layer is used after the batch normalization layer.

Max-pooling layer

Spatial features are large. This layer helps to reduce the feature map size and to remove redundancy from the spatial information. Sample reduction helps to reduce the computation cost and to move meaningful information into the feature maps.

Fully connected layer

These layers are described by their name. All the previous layers are connected here. At this stage, all the previous learning layers are merged. The final layer provides the classification of the model. The number of outputs equals the number of classes identified in the image dataset. In this study, there are four classes.

Softmax layer

The output obtained from the fully connected layer is not normalized. Normalization is performed using an output Softmax layer. The output obtained from this layer is a positive integer and can be used for classification.

Classification layer

The classification layer is the last in the CNN architecture. The Softmax layer uses it for classification. Probabilities are returned for each input image to authenticate the manually labelled classes. It also calculates the loss values.

Figure 2 shows the complete generalized architecture for the CNN architecture. In this study, we are using CNN architectures DenseNet169 and ResNet50. The DenseNet169 architecture is based on the CNN architecture divided into various dense blocks.

We discussed that the DenseNet-169 and ResNet-50 are based on the complete CNN architecture. Because these algorithms are the updated versions of the CNN architecture, some modification is performed in these two algorithms, specifically in object detection from images.

A dense convolutional network (DenseNet-169) is a collection of DenseBlocks with many layers. Densenet1-69 has [6, 12, 32, 32] layered structures, but all the basic structures are the same as the CNN architecture. It has a different layered structure. DenseNet-169 can object detection and diagnosis. Another best feature of the DenseNet-169 is the ability to train the dataset more efficiently and use shorter connections between the layers (Nawaz et al., 2021).

Resnet-50 is also the derived version of CNN architecture, but it is more related to the DenseNet structure. ResNet-50 has a total of 48 convolutional layers. But it has 1 MaxPool layer and 1 Average Pool layer. ResNet-50 has other features, like it has 3.8 × 10⁹ operations of floating. ResNet-50 is the best fit to train the dataset for object detection and diagnosis in the images (Budhiraja & Garg, 2021). The ResNet-50 learning block is shown in Fig. 3.

Experimental results and discussion

This section elaborates on the dataset used for training and testing the model, the experimental analysis for the proposed model, and the outcomes.

Results using DenseNet-169 and ResNet-50

Table 1 shows the statistical results of the training and testing after the successful execution of the model. DenseNet169 returned better values in the training and testing of the model in comparison with the ResNet50 architecture. Based on this, we decided to use the DenseNet-169 model for further processing. Results and visuals are shown in Fig. 3 with the model loss and model AUC values 10.

Table 1:

Evaluation of training and testing model.

No.	Algorithm	Training AUC	Testing AUC	Model loss
1	DenseNet-169	0.9777	0.8870	0.3425
2	ResNet-50	0.8382	0.8198	0.9301

DOI: 10.7717/peerj-cs.1177/table-1

The training results can be seen in Fig. 3 for the DenseNet169 architecture in the form of model loss and model AUC graphs. For the training and validation, DensNet-169 outperforms, as shown in Fig. 4.

After successfully training the proposed model using the training dataset, the model was evaluated in the test phase, where four test cases were performed sequentially. Figure 5 shows test case 1, where the model successfully classified an image labelled “healthy brain” with a probability of 99.93%. Here we select the trained model and pass the image from the test folder. The classification is the output with the best probability value.

Figure 6 shows test case 4, where classification is performed for the “Very Mild AD” class and the proposed model is classified accurately with 91.97% probability.

Figure 7 shows test case 2 for testing the Mild AD class and the trained model classified it successfully. The probability of the identified classification was 97.86%.

Figure 8 shows the test case 3 results for the Moderate AD test data and the proposed model successfully predicted the classification with 86.79% probability.

After successfully training the proposed model using the DenseNet-169 and ResNet-50 CNN architectures, testing proved that the proposed model outperformed the diagnosis and classification of AD. This application could be used in the future to aid in the real-time diagnosis and classification of medical images.

Conclusion and future work

It has been determined that Alzheimer’s disease is an incurable neurodegenerative disease that affects brain memory, particularly in the elderly. Owing to the enormous number of patients, it is impossible to perform manual diagnosis efficiently and health specialists make errors during evaluation due to time constraints and the difficulty of the process. Various procedures are used to diagnose and characterize Alzheimer’s, but an accurate and timely diagnostic solution is required. The proposed model suggests a deep learning-based method for diagnosing and classifying Alzheimer’s disease utilizing the DenseNet-169 and ResNet-50 CNN architectures. Non-Dementia, Very Mild-Dementia, Mild Dementia, and Moderate Dementia were the four classifications of Alzheimer’s Disease in this model. During the training and testing stages, the DenseNet-169 method outperformed. This suggested approach may be used to do real-time analysis and classification of Alzheimer’s disease. In the future, we plan to extend the disease detection with more data sets and use the different measures to detect the system’s accuracy.

Supplemental Information