Alzheimer's disease classification by supervised and intelligent techniques

Abstract

Background

Significant advancements in neuroimaging have emerged over the past decade, notably through positron emission tomography (PET) and magnetic resonance imaging (MRI) for diagnosing Alzheimer's disease (AD) and its precursor, mild cognitive impairment (MCI). Combining imaging modalities with machine learning (ML) techniques enhances diagnostic accuracy.

Objective

To develop predictive models using pre-treatment brain imaging data to distinguish between normal controls (NC), MCI, and AD stages, improving diagnostic precision.

Methods

We utilized the Alzheimer's Disease Neuroimaging Initiative database, processing 3D MRI, PET Florbetaben, and PET Flortaucipir images. Techniques included convolutional neural networks (CNN), fuzzy logic, and multi-layer perceptron (MLP). Feature extraction involved amyloid-β volume, tau protein levels, and empty space volumes.

Results

The fuzzy logic approach achieved a classification accuracy of 99.1%, outperforming CNN (90.67%) and MLP (94%). Integration of multimodal data significantly enhanced performance compared to single-modality approaches.

Conclusions

Our study demonstrates that integrating advanced ML techniques with multimodal neuroimaging can effectively classify AD stages. These findings address critical gaps in early detection and provide a foundation for future clinical applications.

Keywords

Alzheimer's disease classification convolutional neural network fuzzy logic multi-layer perceptron

Introduction

Lately, there has been a surge in the use of artificial intelligence techniques for categorizing neuroimaging data linked to Alzheimer's disease (AD), resulting in substantial advancements. The accurate diagnosis of AD poses a significant challenge in clinical practice due to lack of specificity, especially when assessing the mental status of individuals with impaired consciousness. Furthermore, the distinction between AD and other forms of dementia poses another obstacle. Recently, mild cognitive impairment (MCI), considered a precursor to AD, has attracted attention from the research community due to its association with clinical trials. While MCI may not substantially disrupt daily activities, there is consistent evidence indicating that MCI patients are at a heightened risk of progressing to AD.¹ Neuroimaging methods, including magnetic resonance imaging (MRI)^2,3 and positron emission tomography (PET),^4,5 have been extensively employed in AD assessment, alongside numerous non-imaging biomarkers.⁶ Global data indicates that approximately 35.6 million individuals worldwide are afflicted with AD, with over 7.7 million new cases emerging annually. According to projections by the World Health Organization (WHO), this figure is anticipated to double every 20 years, reaching 65.7 million by 2030 and 115.4 million by 2050. Meanwhile, the Alzheimer's Association of Tunisia reports that over 57,000 individuals in the country currently suffer from the condition. It is estimated that between 4000 and 9000 Tunisian Dinar will be allocated annually for the diagnosis and treatment of each patient. Numerous machine learning techniques have been suggested to assist in AD diagnosis. Biomedical images have commonly been represented using pre-computed medical descriptors. Measurements like volume⁷ and cerebral metabolic rate of glucose⁸ are typically derived from segmented 3D brain regions of interest (ROI) and are utilized for AD classification with methods such as support vector machine (SVM),⁹ Bayesian approaches,¹⁰ or alternative methods.^11,12 Our study presents three approaches for clustering the AD degree that will be developed in the rest of the paper, using PET images, for three binary classification tasks: AD versus normal controls (NC), progressive MCI (pMCI) versus NC, and stable MCI (sMCI) versus NC. These approaches aim to address several key gaps identified in recent studies on AD diagnosis using deep learning and neuroimaging data. The Kang et al.¹³ paper highlights limitations like limited training data, underutilization of 3D spatial information, high computational costs of 3D models, challenges in cross-modality transfer learning, and need for improved feature extraction and model interpretability. The Illakiya and Karthik¹⁴ study, while achieving high classification accuracy, identifies gaps such as limited exploration of interpretability issues, lack of alternative training strategies, limited assessment of generalizability across datasets, lack of detailed architecture design rationale, and need for investigating scalability and computational efficiency. Additionally, the AHANet paper by Illakiya et al.¹⁵ aims to tackle challenges in capturing distinct MCI and AD features, effective integration of global and local features via attention mechanisms, and enhancing multi-scale feature learning for precise classification.

In reality, the various medical specialties are dealing with instability and growing competition from the technology environment. The identification of degenerative disorders is an important task in this field.

Methods

Data used in this study were obtained from the Alzheimer's Disease Neuroimaging Initiative (ADNI) database (http://adni.loni.usc.edu). The ADNI was launched in 2003 as a public-private partnership, led by Principal Investigator Michael W. Weiner, MD. Its goal was to test whether serial MRI, PET, other biological markers, and clinical and neuropsychological assessment can be combined to measure the progression of MCI and early AD. The current goals include validating biomarkers for clinical trials, improving the generalizability of ADNI data by increasing diversity in the participant cohort, and providing data to the scientific community. For up-to-date information, see http://adni.loni.usc.edu.

The most prevalent neurodegenerative illness causing dementia in the older population of developed nations is AD. The number of individuals impacted by AD stands at approximately fifty million, with projections indicating a tripling of this figure by 2050, primarily attributable to population aging.¹⁶

AD manifests as a progressive and irreversible neurological decline, culminating in cognitive function deterioration and eventual patient demise.¹⁷ MCI denotes an intermediate pathological state characterized by diverse symptoms in patients. While MCI may signify the prodromal phase of AD, it can also transition to other forms of dementia.¹⁸ The diagnosis of AD is particularly challenging due to the variety of cognitive and behavioral symptoms that individuals may exhibit. Moreover, the ways in which the disease progresses are just as subjective as the ways in which treatment works. The most difficult objective in this framework is to create novel diagnostic instruments that will aid in the early detection of diseases. For conditions like MCI, computer-aided diagnosis technologies are favored to improve prediction accuracy and support the neuropsychological evaluations carried out by knowledgeable clinicians. To date, numerous investigations in the literature have examined the capacity of machine learning (ML) techniques utilized in the realm of neurodegenerative disorders, including AD has spurred considerable interest in utilizing data from MRI^19,20 or PET.^21,22 Over the past fifteen years, researchers have increasingly focused on diffusion tensor imaging (DTI). DTI, a non-invasive technique, offers insights into the integrity of white matter, which is closely linked to neuropathological processes. It analyzes water diffusion at the microstructural level of the brain, identifying abnormal diffusion patterns in various neurological and neuropsychiatric disorders, including AD.²³ Through DTI, the integrity and trajectory of the major white matter (WM) fiber bundles in the brain can be assessed by following the highly anisotropic diffusion of water along axons. In WM, diffusion is highly anisotropic and less constrained along the axon; in gray matter (GM), on the other hand, diffusion is typically less anisotropic and unrestrained in both directions (isotropic). Based on this supposition, an ellipsoid model has been used to represent the diffusion process, with the lengths of its three primary axes (λ1, λ2, and λ3; shown in Figure 1) corresponding to the diffusion tendencies along each direction.²⁴ DTI is the only neuroimaging modality that can both characterize WM fiber pathways and identify small amounts of WM damage inside these bundles. As such, it can identify deficiencies in anatomical connection that traditional anatomical MRI is unable to identify.^25,26 In the study by Krashenyi et al.,²⁶ to accurately categorize AD, the researchers used a brain MRI image with stochastic properties and an interface system that interfered with logic.

Figure 1.

Color scaled Florbetaben image.

They used an ADNI public base to simulate this system. This system's receiver operating characteristic (ROC) curve yielded an area under curve (AUC) of 0.895 (89.5%).

Classification parameters

Calculation of amyloid-β volume

This value is obtained by going through specific steps. First, we inject a radioactive product called Flore18 into the patient. After an hour, we take a three-dimensional brain image PET called Florbetaben. Then we put the image on a color scale. Then, we calculate the number of yellow and red voxels as expressed in Figure 1. Then at the end, we multiply the number of voxels by the volume of a single voxel ( $0, 2 * 0, 2 * 0, 2 m m 3$ ) and convert it to $c m 3$ .

v o l u m e A m y l o i d = n u m b e r V o x e l A m y l o i d * 0, 2 * 0, 2 * 0, 2

Calculation of tau protein. This value is derived through a series of defined procedures. Initially, a radioactive substance known as Flortaucipir is administered to the patient. Following a one-hour interval, a three-dimensional brain image called Flortaucipir PET is acquired. Subsequently, the image is visualized using a color scale. Then, the count of yellow and red voxels, as delineated in Figure 2, is computed. Finally, this count is multiplied by the volume of a single voxel (0.2 * 0.2 * 0.2 mm) and converted to cubic centimeters (cm³). $v o l u m e T a u = n u m b e r V o x e l T a u * 0, 2 * 0, 2 * 0, 2$

Figure 2.

Color scaled Flortaucipir image.

Figure 3.

MRI structural image.

Figure 5.

Linguistic variable amyloid-β volume.

Calculation of empty space volume

An MRI structural image (Figure 3) that is three-dimensional is used to determine the amount of free space. First, the contour of the skull is drawn so as to omit the void that corresponds to the exterior. Next, multiply the number of black voxels by the voxel volume (0.2*0.2*0.2 mm³) and convert the result to cm³ to find the volume of vacant space. The human brain has an average volume of around. Every ten years, the volume of the brain shrinks by 2% in healthy people and by % in Alzheimer's patients. As a result, the volume of void varies over time within the interval [0, 129]. $v o l E m p t y S p a c e = n u m V o x e l B l a c k * 0, 2 * 0, 2 * 0, 2$

Our approaches

Fuzzy logic classification

The system proposed herein adopts a “Mamdani” type structure, comprising 3 inputs variables: amyloid-β volume, tau protein volume, and empty space volume, along with an output linguistic variable diagnostics (Figure 4) called diagnostics β. The justification for selecting the above inputs is that they have been shown in modern literature on the disease under investigation, AD, to be associated with it. Amyloid-β and the tau protein volumes are the two most well-known and utilized biomarkers of the AD.

Empty space volume

Areas of neuron loss has always correlated with the pathophysiological progression of the disease. These rubrics were selected in consultation with physicians and reviewed literature in the field of medicine-neurology.

Membership function

For defuzzification, we opted for the maximum mean method due to its close alignment with the doctor's diagnosis. This choice was validated through expert feedback, ensuring that the decision-making process mirrors real-world medical interpretations. Additionally, the linguistic variable amyloid-β volume is defined over the interval [0.30] (Figure 5).

In the case of the tau volume variable, its range extends from 0.15. It is categorized into three subsets characterized by triangular membership functions: small, medium, and great, illustrated in Figure 6.

Figure 6.

Linguistic variable tau volume.

The volumeEmptySpace linguistic variable spans across the interval [0.129]. It is divided into three subsets defined by triangular membership functions: small, medium, and great, as depicted in Figure 7.

Figure 7.

The linguistic variable empty space volume.

Figure 4.

Linguistic variable diagnostics.

Figure 8.

Our Neural Network architecture.

Validation and stability of the model

The fuzzy logic approach has been extensively validated using real patient data obtained from publicly available datasets. Furthermore, the convergence properties of this system were tested through repeated iterations, ensuring that the fuzzy classification system reaches stable outputs for a variety of input values. The convergence is particularly influenced by the shape and overlap of the membership functions, which were refined to reduce ambiguity in the diagnosis. Additionally, a stability analysis was performed by varying the input features within a controlled range to ensure consistent diagnostic results.

MLP classification method

Our MLP network consists of three layers: three nodes (neurons) in the input layer; ten nodes with “tansig” nonlinear activation functions in the hidden layer; and three nodes with a “SoftMax” activation function in the output layer. For feature selection, we prioritized variables known to have significant biological relevance to AD progression: empty space volume, amyloid-β volume, and tau protein volume. These variables were selected based on a comprehensive review of medical research and further validated by a neurologist.

Evaluation metrics and model selection

To evaluate the performance of our MLP model, we utilized several metrics beyond simple accuracy, including confusion matrices and ROC curves, to better understand the precision, recall, and F1-score of our classifier, which demonstrate the classifier's performance across various thresholds. These metrics provided insights into the robustness and reliability of the proposed model, ensuring its effectiveness compared to traditional methods.

Convergence properties and stability

The convergence behavior of the MLP network was observed to be stable under various configurations, and the number of iterations required to reach the desired mean squared error threshold was consistent across multiple runs. The stability of the network was verified by testing it on noisy and perturbed data, where it showed resilience and reliable performance. The convergence rate was largely influenced by the learning rate and the number of neurons in the hidden layer, as presented in Table 1.

Table 1.

Table of experimental test.

Learning rate	Iteration number	Neuron number of hidden layer	Accuracy
0.1	1000	3	58.3%
0.1	5000	3	43.1%
0.1	1000	6	65.47%
0.1	10,000	3	39.57%
0.1	10,000	10	74.31%
0.01	1000	3	47.17%
0.01	5000	3	87.15%
0.01	1000	6	53.23%
0.01	10,000	3	91.77%
0.01	150,000	6	81.65%
0.01	10,000	15	85.33%
0.01	10,000	10	94.0%

Parameters selection of our network

Table 1 illustrates the choice of data attributed to the experiments. The parameters used in the classification method are the learning rate, number of iterations, and number of neurons in the hidden layer. The ideal and accuracy choice of these parameters will allow us to reach promising classification results.

CNN classification

Introduction and Feature Selection. Since 2005, several approaches and techniques have been developed by researchers to detect characteristics of AD using medical imaging. This subsection aims to present CNN for contributing to predict disease from brain images. A corrective and punctual treatment can be applied between each layer to improve the quality of expected results. Outputs of an analysis layer allow reconstruction of an intermediate image, which will be a basis for the following layer.

CNN architecture should be able to accomplish some requirements such as:

-Possibility to learn complex and varied characteristics and functions in various classification tasks.

-Ability to learn, low level, intermediate, and high level abstractions with few inputs and human.

-Ability to learn from a very large amount of data (the calculation time for the learning phase, should evolve according to the number of input images).

-Adaptability to new areas and new tasks.

The problem of choosing the architecture of a neural network is always painful and difficult to solve in a data mining process. In fact, there are no specific rules that impose the choice and number of layers to use for achieving such task, their organizations or parameters, related to learning step for finding a perfect architecture, but this is based on experimentation by intercalating, removing or changing the order of the layers, (convolution, pooling, rectification, activation, etc.), and/or by modifying the parameters, learning (learning rate, momentum, number of iterations, etc.) and restart the learning of network again and evaluate, the impact of the modifications made on the behavior and the precision of the model. CNN architecture is formed by a stack of independent processing layers that will be described in the following subsection.

Taking into consideration the further development of medical images analysis, we utilized CNNs in order to predict AD from MRI scans. In this case, we selected a hybrid input featuring dosages of beta-amyloid and Tau protein mass. Our decision was based on the performance of classification based on PET approaches, which usually achieve better outcomes than single modality approaches. However, we chose to include multimodal MRI scan images since they offer a more comprehensive representation of the brain's structural changes enabling better detection of the AD-related patterns.

Parameters and convergence analysis

An intense effort of tuning the parameters of the given model was undertaken, which included variations of the learning rates, convolution kernel sizes, and the number of layers, in order to improve the convergence behavior of the CNN model. The following was found: oscillations in the results decreased with decreased learning rates, with 0 tt learning rate being the most stable and more oscillatory with 0.1. We calculated the stability of the models using such metrics as p-value and ROC curves and provided a very strong quantitative proof for the noted trends.

Scalability and robustness

In order to assess the scalability and robustness of our CNN model, we applied it to a more extensive dataset of MRI images comprising images of different qualities and noise levels. Our proposed architecture (Figure 8) was able to adapt to all different conditions without compromising on classification accuracy and speed of convergence. Additionally, the dropout layers employed in the architecture further reduced overfitting in the network thus improving its performance.

Proposed architecture and validation

The final architecture, whereby the Siamese network and VGGNet components were incorporated, was selected following a thorough performance evaluation against that of single modality models. Due to this dual-branch structure, it was possible to investigate two or more feature sets at the same time thereby making the diagnosis of the disease more complete. The classification results were verified by cross-validation approaches and feedback received from medical professionals not involved with the development of the system, which provided a biologically and clinically relevant context for the system outputs.

Convolution layer (CONV)

The convolutional layers are considered as a crucial component of CNN architectures, they bring out the major part, of the heavy and complex computing. A convolution layer $C^{i}$ (layer i of the network) is parameterized, by its number N of convolution maps $M_{j}^{i} (j \in {1, \dots, N})$ , the size of convolution nuclei is $K_{x} \times K_{y}$ (often square), and the diagram, of connection to the preceding layer $L^{i - 1}$ . Each convolution map $M_{j}^{i}$ is the convolution sum of maps of the previous layer $M_{j}^{i - 1}$ by its respective convolution kernel. A $b_{j}^{i}$ bias is then added, and the result becomes a non-linear transfer function expressed by Eq. 1:

f (x) = 1.7159 \tanh (\frac{2}{3} x)

(1)

In the case of a map fully connected to maps of the previous layer, the result is then expressed by Eq. 2:

M_{j}^{i} = f (b_{j}^{i} + \sum_{n = 1}^{N} M_{n}^{i - 1} * K_{n}^{i})

(2)

A patch of locally linked neurons from the preceding layer is used by the convolutional layers to alter the input data. The input layer's neuronal area and the output layer's locally connected weights will be multiplied by a scalar product in this layer.

Pooling layer

Downsampling layers (pooling) typically come after convolutional layers in standard convolutional neural network topologies. These layers enable poor invariance to rotations and translations that may occur at the input and enable downsampling of the resulting maps. One variation of such a layer that has demonstrated some strengths in the past is the Max-Pooling layer. The output of a Max-Pooling layer is given, by the maximum activation value within input layer for different regions having $K_{x} \times K_{y}$ of size that not be overlapped. A bias is applied, just like in a convolutional layer, and the outcome is sent to the previously mentioned transfer function. A bias neuron has an activation function that is always equal to 1. A bias links to neurons in the previous layer similarly to other neurons, and the intermediary weight value acts as a form of threshold. Neurons and biais are organized as non-looped layer structure (feed-forward). The network can therefore simply interpret like an input-output model with weights and thresholds (adjustable) parameters of the model.

Correction layer (rectified linear unit, ReLU)

Improving processing efficiency is often achievable by incorporating an additional layer between processing layers that applies a mathematical function, known as an activation function, to the output signals. Notably, we have:

- The ReLU correction:

f (x) = max (0, x)

This function, also referred to as the non-saturating activation function, enhances the nonlinear characteristics of the decision function and the network as a whole, while preserving the receptive fields of the convolution layer.

- Hyperbolic tangent correction:

f (x) = \tanh (x)

- Saturating hyperbolic tangent correction:

f (x) = | \tanh (x) |

- Sigmoid function correction:

f (x) = (1 + e^{- x})^{- 1}

ReLU operates as a fundamental operation applied per pixel, with the goal of replacing all negative values in the generated feature maps with zero. Its primary objective is to introduce nonlinearity into CNNs, as real-world data typically exhibits nonlinear characteristics. Often, ReLU is favored due to its ability to facilitate rapid neural network learning, while not substantially impacting generalization accuracy.

Our proposed architecture

The proposed CNN is composed of two concatenated network models and has two entries as shown in Figure 9. Our CNN receives as input two 2D MRI images in gray level with 256 × 256 size and outputs the membership degree of the three basic classes (NC, MCI, AD) of the disease. The architecture of CNN consists of 26 layers. It is defined in two parts: 1) a Siamois network and 2) the continuation of the VGGNet network architecture.

Figure 9.

Architecture of proposed CNN neural network.

The first branch of the Siamois network presents a first convolution layer constituted by 3 × 3 filter and provides in outputs 50 images of 256 × 256 size. This layer is followed by a down sampling layer which uses a Max- Pooling with a 2 × 2 mask, that reduces the size of image to half size (256/2 = 128). Thus, a new convolutional layer is considered, its kernel has

3 × 3 size and 50 filters, this layer followed by down sampling layer of Max-Pooling with a 3 × 3 kernel. Then we have third feature extraction layer with 3 × 3 kernel and a down sampling layer with a 2 × 2 sized. in the second branch, the first convolution layer comprising a filter of size 5 × 5. This layer is followed by a sub-sampling layer which keeps the same Max-Pooling filter of size 2 × 2 as the first branch.

The two outputs of the Samois networks are concatenated followed by a convolution layer with 3 × 3 kernel and 100 filters, followed by a down sampling layer of Max-Pooling with 2 × 2 kernel size, then we have a Flatten layer, fully connected layer and Dropout loss layer that is used to set the fraction rate of the input units to 0 on each upgrade by randomly way through the learning time, that's contributes to avoid over-learning phenomena (Overfitting).

Finally, the outputs are served to a fully connected FC layer which has 3 neurons (which characterize the three states (NC, MCI, AD) of classification) that are fully connected to the previous layers. To improve the efficiency of the treatment, a rectification layer (ReLU) is associated after each convolution layer.

Results

In the current literature, PET-based approaches have provided a suitable way to classify AD. However, our multimodal MRI based methodology has equivalent efficiency while decreasing the need to use specific imaging options and hence is fits different clinical scenarios. According to Table 2, there is a contrast between our model and the more recent research surrounding the use of PET imaging in regard to the use of our model's accuracy standard.

Table 2.

Impact of imaging modalities on classification accuracy.

Modality	Accuracy
3D MRI only	83.2%
3D PET Florbetaben only	91.5%
3D PET Flortaucipir only	86.7%
Combined 3D MRI + PET images	99.1%

Table 12.

Effect of preprocessing techniques on model performance.

Preprocessing Method	Accuracy
No preprocessing	87.3%
Normalization only	90.1%
Noise reduction only	91.0%
Both normalization & noise reduction	99.1%

Future work and scalability

In order to strengthen the evidence for our model, we intend to increase the dataset once more and examine whether the addition of PET images with MRI will work. Such a fusion approach will help improve the final system's predictive performance thereby reducing the effect of classification based on one image type. Finally, more quantitative analysis will be performed to evaluate how well our model can be generalized in terms of patients and imaging center to expand its usability in practice.

Dataset

Our research leverages 3D MRI, 3D PET Florbetaben, and 3D PET Flortaucipir medical images sourced from the ADNI public database. Using data from 173 individuals, we employed these imaging modalities to gain a comprehensive understanding of AD pathology. Below, we provide a detailed description of each imaging modality and its relevance to AD research.

Ablation study

This section's goal is to examine how different aspects of the suggested approach affect its categorization performance. Here, we assess the effects of model parameters, preprocessing procedures, and several imaging modalities (3D MRI, 3D PET Florbetaben, and 3D PET Flortaucipir) on the overall classification accuracy. The same dataset, consisting of 173 individuals, is used for each experiment, and the outcomes are compared to determine the most important factors for attaining high performance.

Impact of imaging modalities

We first examine the individual contributions of the three imaging modalities—3D MRI, 3D PET Florbetaben, and 3D PET Flortaucipir—in the classification process.

3D MRI scans are essential for evaluating brain atrophy, especially in areas like the hippocampus, and provide structural information. However, biochemical indicators like tau tangles and amyloid plaques are not seen by MRI alone.

3D PET Florbetaben has a major benefit in terms of illness identification since it is made to view amyloid plaques, which are indicative of AD. Accuracy is much increased when all three modalities are used together, demonstrating that combining anatomical and biochemical data is more efficient than depending just on one modality.

Effect of preprocessing techniques

We also investigate the effect of different preprocessing techniques on classification accuracy. These include normalization, noise reduction, and spatial alignment of the images.

Because of the inherent noise and inconsistencies in the imaging data, raw data without preprocessing produces less-than-ideal results. Normalization and noise reduction are essential for enhancing the data quality, which has a major impact on the model's performance.

Impact of model configuration

We experiment with several CNN and MLP architectures to evaluate the effect of the model's configuration. There are variations in the number of layers, neurons per layer, and kernel sizes.

When the number of hidden nodes is increased, the MLP architecture performs well; with 10 nodes in the hidden layer, it achieves 94% accuracy. CNNs produce competitive results; somewhat higher accuracy is obtained with combinations of small (3 × 3) and big (5 × 5) convolutional kernels and several layers.

Summary of ablation study results

3D PET With a classification accuracy of 99.1% (Table 3), the best classification results are obtained when MRI scans and Florbetaben are coupled.

The model's performance is greatly improved by preprocessing methods like noise reduction and normalization.

A hidden layer with 10 nodes works best for the MLP model, whereas a mix of 3 × 3 and 5 × 5 kernels with 3 layers produces the most dependable results for the CNN.

The CNN model's training procedure is optimized using a learning rate of 0.01 and 50 epochs.

3D MRI

MRI provides high-resolution anatomical images of the brain, allowing for the assessment of structural changes associated with AD. The 3D MRI images are used to measure brain atrophy, particularly in regions vulnerable to AD such as the hippocampus and entorhinal cortex.

Detailed anatomical visualization

MRI offers superior spatial resolution, enabling precise measurement of brain structures.

Quantitative metrics

Volume measurements and cortical thickness analyses help quantify the extent of brain atrophy.

Differential diagnosis

MRI is essential for distinguishing AD from other neurological conditions that may present with similar clinical symptoms but have different structural brain changes.

3D PET Florbetaben

Florbetaben is a radiotracer used in PET imaging to visualize amyloid-β plaques in the brain, which are a hallmark of AD. The 3D PET Florbetaben images allow for the detailed examination of amyloid plaque deposition in various brain regions, facilitating early diagnosis and monitoring of disease progression.

Comparison with existing systems

Our suggested strategy, which combines multimodal imaging (MRI + PET) with sophisticated preprocessing techniques, exhibits notable accuracy gains over current approaches. A comparison of our approach with a number of current systems may be found below:

By using a fuzzy logic-based approach (Table 4), our method achieves an accuracy of 99.1%, which is significantly better than the methods of Liu et al. (53.8%)¹¹ and Hosseini-Asl et al. (89.1%).²² A major factor in this performance improvement is the combination of multimodal data (MRI and PET) with preprocessing methods like noise reduction and normalization. This comparison demonstrates how our approach is more robust than previous systems in the literature by integrating preprocessing approaches, sophisticated model architectures, and multimodal imaging.

We used 3D PET Florbetaben 3D, 3D Flortaucipir PET, and 3D MRI medical pictures from the ADNI public database in our research. Matlab2017a was used for our analysis. Because the Mamdani method is close to clinical diagnosis and appropriate for human intervention, we used it. The two classes NC and AD were perfectly classified using the suggested method. In contrast to the other classes, the MCI class had a minor decline in recognition rates. 99.1% is the average categorization rate.

Table 5 offers a comparative comparison of different approaches according to the three classes’ accuracy (NC, MCI, and AD). Notably, with an identification rate of 99.1%, the logic-based approach showed the highest performance.

For MLP technique, we used the Matlab (Pattern Recognition Network) toolbox.

Using data from 173 individuals, we utilized 3D MRI medical imaging, 3D PET Florbetaben 3D, and 3D TEP Flortaucipir that were obtained from the ADNI public database. We trained the system across 10,000 epochs with a learning rate of 0.01. Actually, we ran a lot of experiments by changing the number of epochs. For example, the loss rate becomes unstable again when the number of epochs surpasses 10,000; at that point, the epoch with the best loss rate is 10,000. The suggested method enhances an error-free categorization for the two classes NC and AD. For the MCI class the recognition rate is reduced compared to others as expressed in Table 6 of the confusion matrix. The rate average classification is 94%.

Table 6.

Impact of model configuration on classification accuracy.

Model Configuration	Accuracy
MLP with 3 layers, 3 nodes	83.4%
MLP with 3 layers, 10 nodes	94%
CNN with 2 layers, 3 × 3 kernel	88.39%
CNN with 3 layers, 3 × 3 + 5 × 5 kernels	90.67%

Table 3.

Summary of ablation study results.

Approach	Accuracy
Liu et al.¹¹	53.8%
Hosseini-Asl et al.²²	89.1%
Proposed Method (Fuzzy Logic)	99.1%
Proposed Method (MLP)	94%
Proposed Method (CNN)	90.67%

Table 5.

Table type styles.

Approach	Accuracy
Liu et al.¹¹	53.8%
Hosseini-Asl et al.²²	89.1%
Proposed method (fuzzy logic)	99.1%

Table 4.

Approach statistics table.

Approach	Accuracy
Liu et al.¹¹	53.8%
Hosseini-Asl et al.²²	89.1%
Proposed method (MLP)	94%

Concerning CNN technique, Figure 10 shows the error curve according number of epochs. Indeed, when the number of epochs increases, rate of loss decreases. It has been seen from Figure 2.15 that the loss curve converges to 0 and remains constant up to epoch 50, which leads to the most suitable number of epochs being 50 as shown in Table 7.

Figure 10.

Loss curve according epochs number.

Table 7.

Confusion matrix.

Output classTarget class	NC	MCI	AD
NC	100.0%	0.0%	0.0%
MCI	11.3%	82.0%	6.7%
AD	0.0%	0.0%	100.0%

Table 9.

Confusion matrix.

Parameters	Value
Learning Rate	0.01
Epochs Number	50

We carried out several tests by varying the number of epochs such that when the number of iterations exceeds 50 the loss rate is disturbed again, it is the overfitting then the best rate of loss is for the epoch 50.

The loss function's convergence during training is depicted in the plot. Since the model begins with random weights, the initial loss is substantial, as is common. Rapid learning and better predictions are indicated by the notable drop in loss during the initial epochs. The loss curve flattens after this first decline, indicating that the model has largely absorbed the patterns in the data and that more training or hyperparameter tweaks will be necessary for future gains. In the end, the modest and steady loss suggests that the model has reduced the error, and additional epochs are unlikely to produce appreciable gains unless overfitting takes place. The model learns effectively at first and stabilizes as it gets closer to the minimum loss, exhibiting successful convergence.

Several experiments and tests were carried out as shown in Table 8.

Table 8.

Table of experiments and tests.

Learning rate	Iteration	Convolution S1	Convolution S2	Accuracy
0.1	10	5*5	5*5	21.45%
0.1	10	3*3	3*3	34.42%
0.01	10	5*5	5*5	23.51%
0.01	10	3*3	3*3	27.81%
0.01	20	3*3	3*3	41.23%
0.01	50	3*3	3*3	65.31%
0.01	50	5*5	5*5	61.17%
0.01	100	3*3	3*3	74.69%
0.01	100	5*5	5*5	70.64%
0.01	100	3*3	5*5	88.39%
0.01	60	3*3	5*5	89.55%
0.01	50	3*3	5*5	90.67%

The parameters used in the classification method are learning rate (Table 9), number of iterations, and the convolution kernel dimensions. The ideal and correct choice of these parameters will help and ensure more precise classification results. In this approach, we used 2D MRI images of the hippocampus brain of 394 people presented in Table 10. 65% of this data base is devoted to learning step and 35% is devoted to testing step.

Table 10.

Demographic information of patients.

	CN	MCI	AD
Number M/F	39/41	89/75	74/76
Age M/F	79/80	79/80	78/85

As shown in Table 11, the confusion Matrix proves the perfect classification for classes NC and AD. For the MCI class, the recognition rate has experienced a decline in comparison to the others. The average classification rate stands at 90.67%.

Table 11.

Confusion matrix.

Output classTarget class	NC	MCI	AD
NC	100.0%	0.0%	0.0%
MCI	11.3%	82.0%	6.7%
AD	0.0%	0.0%	100.0%

The fuzzy logic approach adopts the Mamdani logical system, amalgamating a language output variable with three linguistic input variables: beta-amyloid volume, Tau protein volume, and empty space volume. Our approach employs standard defuzzification logic and functions, accommodating up to 8 rules. This methodology leverages 3D MRI, 3D PET Florbetaben, and 3D PET Flortaucipir images, achieving a classification rate of 99.1%. For the MLP technique, our network comprises three layers: an input layer with three nodes, a hidden layer with ten nodes, and an output layer with three nodes. This technique uses the same data base than fuzzy logic method, we have reached 94% as a rate of classification. For the CNN technique, many tests have been made according to the main parameters, like learning rate, number of epoch, and mask with various size for the two branch used, the best parameters selected are respectively 0.01, 50, 3*3 and 5*5, the classification rate reached is 90.67.

Discussion

Our results demonstrate the efficacy of integrating multimodal imaging and advanced machine learning techniques in AD classification. The fuzzy logic approach yielded superior accuracy (99.1%), attributed to its ability to combine anatomical and biochemical markers effectively. Compared to prior studies, our methods outperformed approaches reliant on single modalities or conventional classifiers.

The robustness of preprocessing methods like normalization and noise reduction was evident, significantly enhancing model performance (Table 12). Additionally, parameter tuning in CNN and MLP architectures proved critical in optimizing classification accuracy.

Footnotes

Acknowledgments

Data collection and sharing for the Alzheimer's Disease Neuroimaging Initiative (ADNI) is funded by the National Institute on Aging (National Institutes of Health Grant U19AG024904). The grantee organization is the Northern California Institute for Research and Education. In the past, ADNI has also received funding from the National Institute of Biomedical Imaging and Bioengineering, the Canadian Institutes of Health Research, and private sector contributions through the Foundation for the National Institutes of Health (FNIH) including generous contributions from the following: AbbVie, Alzheimer's Association; Alzheimer's Drug Discovery Foundation; Araclon Biotech; BioClinica, Inc.; Biogen; Bristol-Myers Squibb Company; CereSpir, Inc.; Cogstate; Eisai Inc.; Elan Pharmaceuticals, Inc.; Eli Lilly and Company; EuroImmun; F. Hoffmann-La Roche Ltd and its affiliated company Genentech, Inc.; Fujirebio; GE Healthcare; IXICO Ltd; Janssen Alzheimer Immunotherapy Research & Development, LLC.; Johnson & Johnson Pharmaceutical Research & Development LLC.; Lumosity; Lundbeck; Merck & Co., Inc.; Meso Scale Diagnostics, LLC.; NeuroRx Research; Neurotrack Technologies; Novartis Pharmaceuticals Corporation; Pfizer Inc.; Piramal Imaging; Servier; Takeda Pharmaceutical Company; and Transition Therapeutics.

ORCID iD

Jabli Mohamed Amine

Author contributions

Mohamed Amine Jabli (Methodology; Writing – original draft; Writing – review & editing); Mousa Mourad (Supervision).

Funding

The authors received no financial support for the research, authorship, and/or publication of this article.

Declaration of conflicting interests

The authors declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.

Data availability

The data supporting this study are available from the ADNI database and can be accessed through their public repository.

References

Dubois

Feldman

Jacova

, et al. Research criteria for the diagnosis of Alzheimer’s disease: revising the NINCDSADRDA criteria. Lancet Neurol 2007; 6: 734–746.

Davatzikos

Bhatt

Shaw

, et al. Prediction of MCI to AD conversion via MRI, CSF biomarkers, and pattern classification. Neurobiol Aging 2011; 32: 2322.e19–e27.

Cai

Liu

Song

, et al. A 3D difference of Gaussian-based lesion detector for brain PET. In: 2014 IEEE 11th International Symposium on Biomedical Imaging, ISBI 2014, Beijing, China, 29 April 2014-2 May 2014, pp.677–680, Piscataway, NJ: Institute of Electrical and Electronics Engineers (IEEE).

Chetelat

Desgranges

, , et al.

Mild cognitive impairment: can FDG-PET predict who is to rapidly convert to Alzheimer’s disease?

Neurology 2003; 60: 1374–1377.

Liu

Cai

Wen

, et al. Multi-channel neurodegenerative pattern analysis and its application in Alzheimer’s disease characterization. Comput Med Imaging Graph 2014; 38: 436–444.

Bouwman

Schoonenboom

SNM

van der Flier

, et al. CSF Biomarkers and medial temporal lobe atrophy predict dementia in mild cognitive impairment. Neurobiol Aging 2007; 28: 1070–1074.

Risacher

Saykin

West

, et al. Baseline MRI predictors of conversion from MCI to probable AD in the ADNI cohort. Curr Alzheimers Res 2009; 6: 347–361.

Cai

Liu

Wen

, et al. 3D Neurological image retrieval with localized pathology-centric CMRGlc patterns. In: 2010 17th IEEE International Conference on Image Processing, ICIP 2010, Hong Kong, 26 September 2010-29 September 2010, pp.3201–3204, Piscataway, NJ

Institute of Electrical and Electronics Engineers (IEEE).

Zhang

Wang

Zhou

, et al. Multi-modal classification of Alzheimer’s disease and mild cognitive impairment. Neuroimage 2011; 55: 856–867.

10.

Liu

Song

Cai

, et al. Multifold Bayesian kernelization in Alzheimer’s diagnosis. Med Image Comput Comput Assist Interv 2013; 16: 303–310.

11.

Liu

Cai

Wen

, et al. Semantic-word-based image retrieval for neurodegenerative disorders. J Nucl Med 2012; 53: 2309.

12.

Zhang

Song

Liu

, et al. Semantic association for neuroimaging classification of PET images. J Nucl Med 2014; 55: 2029.

13.

Kang

Lin

Sun

, et al. Three-round learning strategy based on 3D deep convolutional GANs for Alzheimer’s disease staging. Sci Rep 2023; 13: 5750.

14.

Illakiya

Karthik

. A dimension-centric proximate attention network and Swin transformer for age-based classification of mild cognitive impairment from brain MRI. IEEE Access 2023; 11: 128018–128031.

15.

Illakiya

Ramamurthy

Siddharth

, et al. AHANet: adaptive hybrid attention network for Alzheimer’s disease classification using brain magnetic resonance imaging. Bioengineering 2023; 10: 714.

16.

Ouchtati

Chergui

Mavromatis

, et al. Novel method for brain tumor classification based on use of image entropy and seven Hu’s invariant moments. Traitement Signal 2019; 36: 483–491.

17.

Tingmei

Yanyan

. A nonlinear tensor-based machine learning algorithm for image classification. Rev Intell Artif 2019; 33: 475–481.

18.

Iman

Hasan

Hiroshi

. Classification of Alzheimer’s disease and prediction of mild cognitive impairment-to-Alzheimer’s conversion from structural magnetic resonance imaging using feature ranking and a genetic algorithm. Comput Biol Med 2017; 83: 109–119.

19.

Hajer

Rostom

Nawres

. Machine learning and deep learning for clinical data and PET/SPECT imaging in Parkinson's disease: a review. IET Image Process 2020; 14: 4013–4026.

20.

Suk

H-I

Lee

S-W

Shen

, et al. Hierarchical feature representation and multimodal fusion with deep learning for AD/MCI diagnosis. Neuroimage 2014; 101: 569–582.

21.

Liu

Zhang

Shen

, et al. Multimodal neuroimaging feature learning for multiclass diagnosis of Alzheimer's disease. IEEE Trans Biomed Eng 2015; 62: 1132–1140.

22.

Hosseini-Asl

Keynton

El-Baz

. Alzheimer's disease diagnostics by adaptation of 3D convolutional network. In: Proceedings of the Medical Image Computing and Computer-Assisted Intervention, Munich, Germany, 5–9 October 2016, pp.238–247, Cham, Switzerland: Springer.

23.

Liu

Cheng

Wang

, et al. Multi-modality cascaded convolutional neural networks for Alzheimer's disease diagnosis. Comput Med Imaging Graph 2018; 70: 101–111.

24.

Zadeh

. Fuzzy sets. Inform Control 1965; 8: 338–353.

25.

Zadeh

. Fuzzy algorithms. Inform Control 1968; 12: 94–102.

26.

Krashenyi

Popov

Ramirez

, et al. Fuzzy classification of Alzheimer's disease using statistical moments. In: 2015 IEEE 35th International Conference on Electronics and Nanotechnology (ELNANO), Kyiv, Ukraine, 21-24 April 2015, pp.409–412, Piscataway, NJ: Institute of Electrical and Electronics Engineers (IEEE).