Sometimes we found signs of AD in the brain data of healthy and older patients, so considerable experience and knowledge were essential to distinguish the AD data from the HC patients’ data. In this context, we have experimented the performance of the proposed CAD system on the OASIS [26] and ADNI [27] datasets.

This work used two datasets (ADNI and OASIS), which are available in the public domain. For the benchmark ADNI dataset, the terms of use are declared on their website [28]. All patients in the ADNI database provided written informed consent, which was approved by the institutional review board of each participating institution. Patients were informed that their information would be kept confidential and their data would be anonymous and would be part of scientific publications.

According to local legislation and institutional requirements, the study of human participants using the OASIS dataset does not require ethical review and approval [26]. Written informed consent from the patients’ legal guardians or next of kin was not required to participate in this study in accordance with national legislation and institutional requirements [26]. The data used for the analysis has been deidentified and made public.

We performed the following steps on the OASIS and ADNI neuroimages: normalization, resizing, removing nonbrain slices, selecting slices with the most information, and converting 3D images into 2D slices. First, the damaged original files containing the images were removed. We selected a larger number of central slices to aid the CAD system in accurately classifying AD. We used an SPM tool (SPM8 [29]), which is a major update to SPM software, originally developed by Karl Friston, to partially correct spatial intensity inhomogeneities. This software normalized all the images using a general affine model with 12 parameters. The origin of the raw sMRI scans was set manually to anterior commissure before manually registering them with SPM’s canonical T1 template image. We applied the nonparametric nonuniform intensity normalization (N3) technique to solve the tissue intensity nonuniformity problem [30]. Then the hybrid median filter was used to remove impulse noise while preserving edges.

We have carried out the performance assessment and the comparison of the classifiers using typical confusion matrix–based evaluation metrics. The confusion matrix has the elements of true positive (TP), false positive (FP), false negative (FN), and true negative (TN). Each column of the matrix indicates an instance of the predicted class, and each row contains a true (correct or actual) class. The following are the metrics used to evaluate the performance of the CAD system.

Sensitivity—also known as recall—is used for calculating the classifier’s ability to correctly predict Alzheimer instances (AD class). On the other hand, the classifier uses specificity to accurately predict all non-Alzheimer instances (HC class) across all inputs.

A classifier should have high sensitivity and specificity. Therefore, the accuracy metric, which calculates the number of correctly classified instances relative to the total number of instances, is the average of these two measures. The precision metric measures the classifier’s ability to quantify the number of TPs of the AD class that receive a correct label in classification.

The combined harmonic mean of both sensitivity and precision gives the F₁-score, which takes a value between 0 and 1. The receiver operating characteristic curve, a method for visualizing a classifier’s ability to diagnose or predict correctly, clearly illustrates the trade-off that arises between the sensitivity and specificity metrics. At various thresholds, the receiver operating characteristic curve plots the TP rate or sensitivity against the FP rate (1 – specificity).

We aim to determine the degree of separability, or the ability to correctly predict class, using the AUROC. The higher the AUROC, the better; 1 would be perfect, and 0.5 would be random. Accuracy, sensitivity, specificity, precision, F₁-score, and AUROC are the six main metrics used to assess the efficacy of each classifier. The following are the mathematical formulas for the first five metrics.

The main finding is that the development of diagnostic tools applying the ML approach in conjunction with neuroimaging data could substantially help in automating the classification and prediction of AD.

In this context, this study proposed a complete CAD system to successfully classify patients with AD and discriminate them from HC patients. The purpose was to examine the association between SVM, FFNN, and ViT ML classifiers; PCA, LDA, and t-SNE dimensionality reduction techniques; and sMRI and FDG-PET neuroimaging modalities to detect early signs of AD. Furthermore, we aimed to clarify the impact of some data preprocessing strategies, such as noise reduction and data augmentation, on improving the performance of classifiers.

With regard to the sMRI and FDG-PET modalities, they can provide large amounts of information; nevertheless, interpreting all image content is challenging for physicians. The experimental analysis demonstrates that combining these neuroimaging modalities with selected ML classifiers enhances their performance, enabling doctors to provide precise diagnosis and timely patient care. This confirms the theory regarding the benefits of these two modalities. Since sMRI provides high-resolution images of brain anatomical structures, which confirm structural change in the brain, it shows shrinkage of brain tissue and abnormalities, while FDG-PET shows the functionality of the brain.

Regarding the selected dimensional reduction techniques, all of the chosen dimensional reduction techniques performed well as feature extractors when combined with the SVM and FFNN classifiers, but a comparative analysis of the three techniques reveals that PCA outperforms LDA and t-SNE. However, it is important to clarify certain findings: PCA allows the identification of the most significant variables in the data due to its potential to generate new variables, which represent linear combinations of the original variables. Moreover, t-SNE differs from PCA by preserving only small pairwise distances or local similarities, while PCA aims to preserve large pairwise distances to maximize variance. Unlike PCA, LDA is a supervised technique that maximizes class separability in the reduced dimensionality space, thereby retaining the most discriminative features.

Preliminary results from evaluating the complete CAD system using the three classifiers prove that the system is more effective in separating AD and HC classes. The results provided by all the experiments carried out reveal an increase in sensitivity and, consequently, the final accuracy obtained by the basic VAF-SVM model (66.3% for OASIS and 42.8% for ADNI). We compared the performance of the SVM, FFNN, and ViT models using confusion matrix–based metrics.

All models performed well, providing acceptable performance for both databases. Data augmentation/ViT outperformed other models, with accuracies of 93.2% for OASIS and 90.4% for ADNI (see Tables 4 and 5 for more details on results obtained from all models tested on both databases). The second best classifier is PCA/SVM, achieving an accuracy decrease of 1.3% for OASIS and 1.8% for ADNI, compared to the rates obtained by ViT, resulting in overall accuracy rates of 91.9% and 88.6% for OASIS and ADNI, respectively. Therefore, the data augmentation process and the PCA dimensionality reduction method have the potential to impact the overall performance of the ViT and SVM models, respectively.

Moreover, compared to the performance using a single MRI modality, all models performed well using a multimodal MRI/PET environment. The best results with MRI were also obtained with ViT and SVM classifiers. Accuracies of 83.9% for the OASIS dataset and 81.2% for ADNI were obtained using the data augmentation/ViT approach. PCA/SVM achieved accuracies of 82.4% for the OASIS and 80.6% for the ADNI datasets. This draws attention to the potential of integrating multiple modalities to increase the performance of the CAD system.

To verify the convergence of the proposed CAD system, we compared the results obtained with some relevant state-of-the-art ML models. The experimental results show that our models, particularly SVM and ViT, have good performance on both the OASIS and ADNI datasets and achieved better or comparable accuracy to most existing methods in the literature. For the OASIS dataset, the PCA/SVM method had a 91.9% accuracy and the ViT model with data augmentation had a 93.2% accuracy. Nanni et al [33], Khan and Zubair [16], Sethi et al [2], Basheer et al [34], Saratxaga et al [35], and Liu et al [36] got 90.2%, 86.8%, 86.2%, 92.3%, 93%, and 82.6% accuracy, respectively.

The same finding was obtained for the ADNI dataset, where we achieved an accuracy of 88.6% using the PCA/SVM approach and 90.4% using the ViT model by increasing the data. In contrast, the accuracy achieved by Rallabandi et al [37], Jo et al [4], Jo et al [3], Liu et al [36], and Shojaei et al [38] was 75%, 75.02%, 80.8%, 90%, and 87%, respectively. Table 6 compares our best results obtained with the prior state-of-the-art models discussed.

There are several improvements possible for the proposed CAD system. We aim to enhance the system’s performance by collaborating with more extensive AD datasets and implementing various types of ANN and ML-based classifiers.

The PCA used for feature extraction looks for the principal axis direction, which is used to effectively represent the common features of similar samples. This is very effective for representing the common features of the same kind of data samples, but it is not suitable for distinguishing different sample classes. Therefore, to achieve the purpose of feature extraction, we need to combine PCA with other feature dimensionality reduction algorithms like uniform manifold approximation and projection.