Alzheimer’s detection using discrete wavelet transform based image fusion and vision information transformer

1 Introduction

Alzheimer’s disease (AD) is a progressive neurological condition that predominantly affects memory, cognition, and behavior. It is the most prevalent cause of dementia, an illness characterised by a significant deterioration in cognitive ability that interferes with everyday life (Amir et al., 2020). Named after Dr. Alois Alzheimer, who first described the condition in 1906, AD gradually worsens over time, eventually robbing individuals of their ability to carry out even the simplest tasks. As of 2015, approximately 5.3 million individuals were estimated to be affected by AD. However, projections suggest a substantial rise in this figure by 2050, with expectations of around 16 million cases. This stark increase underscores the escalating impact of AD on global healthcare systems, highlighting the critical necessity for improved diagnostic methods (Amin Hazarika et al., 2021).

Medical imaging plays a decisive role in the diagnosis and management of AD, giving important insights into the anatomical and functional changes that occur in the brain. Here is an overview of the medical imaging-based diagnosis process for AD:

Magnetic Resonance Imaging (MRI): MRI is a non-invasive imaging technology that combines strong magnets and radio waves to obtain detailed pictures of the brain’s structure (Fayed et al., 2012). In the diagnosis of AD, MRI can reveal structural changes such as shrinkage of the hippocampus and other regions of the brain associated with memory and cognition (Feng et al., 2020). These changes are indicative of neurodegeneration and are commonly observed in individuals with AD (Salvatore et al., 2015). Furthermore, an MRI can help rule out other illnesses that may resemble Alzheimer’s symptoms, such as stroke or brain tumours.

Positron Emission Tomography (PET) Imaging: PET imaging allows for the visualisation of metabolic and molecular processes in the brain. One of the most widely used PET imaging techniques in AD diagnosis is fluorodeoxyglucose (FDG)-PET, which measures glucose metabolism in the brain (Noble and Scarmeas, 2009). In AD, there is a typical pattern of impaired glucose metabolism in distinct brain areas, particularly the posterior cingulate cortex and temporoparietal regions (Noble and Scarmeas, 2009). This pattern of hypometabolism can help distinguish AD from other types of dementia and may aid in early detection (Marcus et al., 2014).

Overall, medical imaging plays a critical role in the diagnosis and management of AD, providing valuable information about brain structure, function, and pathology. By combining MRI and PET imaging, the brain can provide both structural and functional/molecular information, providing a more comprehensive understanding of the neurodegenerative processes underlying AD (Tong et al., 2017; Shi et al., 2017). It can improve the accuracy of AD diagnosis by allowing correlation between structural changes and metabolic and molecular abnormalities in the brain.

During the past decade, artificial intelligence has shown significant utility in the diagnosis of AD. Among the prevailing classification techniques, shallow machine learning (ML), convolutional neural networks (CNN), and deep learning (DL) have emerged as the most widely utilised methods (Al-Naami et al., 2013; Alam et al., 2017a; Alam et al., 2017b; Chaplot et al., 2006). DL has been shown to be beneficial for large image data sets (Basaia et al., 2019; Amir et al., 2020).

Shen et al. (2017) conducted a survey on DL for AD. It can be concluded that ML techniques are useful in determining the underlying neurological causes of brain disorders, but DL methods overcome the necessity of feature engineering (Ma et al., 2018). The need for hand-crafted features in standard ML algorithms is a disadvantage that may result in less-than-ideal performance. However, DL methods are quite successful because they automatically extract significant features from the data.

The growing GPU processing capability enabled the use of DL techniques for image classification applications (Plis et al., 2014). Tanveer et al. (2021) have suggested a deep transfer ensemble, an ensemble of deep neural networks (DNN) trained with transfer learning (TL), for the classification of AD. Lu et al. (2018) suggested a unique multimodal DNN with a multistage method to recognise dementia sufferers. With 82.4% accuracy, this technique can identify who would get mild cognitive impairment (MCI) and who will go on to develop AD 3 years later. For the class of people with AD, the model achieves a sensitivity of 94.23% and an accuracy of 86.3%. In order to diagnose AD, Ahmed et al. (2019) suggested the CNN ensemble model for feature extraction and the SoftMax classifier. Using the left and right hippocampal regions from MRI images, this model avoids overfitting and achieves an accuracy of 90.05%. In Jain et al. (2019) for the extraction of features, a pre-trained VGG16 model was used, together with a FreeSurfer for pre-processing, entropy-based MRI slice selection and TL-based classification, resulting in the mathematical model PFSECTL. The researchers were able to obtain 95.73% precision. Shi et al. (2017) suggested a deep polynomial network that works well for small and large data sets. Using the ADNI dataset, the model’s accuracy for binary and multiclassification is 55.34%. Wang et al. (2019) used a 3D ensemble model of convolutional networks to present AD and MCI. Using a probability-based fusion strategy, 3D-DenseNets were optimised. Using the ADNI dataset, the model achieves a classification accuracy of 97.52%. To learn the characteristics of AD, Ge et al. (2019) proposed a 3D multiscale DL architecture. The system obtained a test accuracy of around 93.53% on a subject-segregated random brain scan partitioned dataset, with an average accuracy of 87.24%. In a DL pipeline proposed by Sarraf et al. (2016), the CNN model is trained using a large number of training images to perform feature categorisation using scale- and shift-invariant methods. For functional magnetic resonance imaging and MRI, the model achieves 94.32% and 97.88%, respectively. On the one hand, DL models have been successfully used to analyse neuroimaging data from MRIs (Lecun et al., 2015). On the other hand, TL has gained popularity, which enables DL training to be effective in the situation of insufficient data (Jadhav et al., 2021).

In recent studies, CNN-based models have been widely used for the detection of AD, leaving a scope for Vision Information Transformer (ViT) based studies for AD detection. ViT offers several advantages over traditional CNNs, particularly in their ability to understand global context, scalability, flexibility, and performance on large datasets (Chen et al., 2024). Dai et al. (2021) proposed a transformer-based multi-modal image analysis model. They integrate the strengths of CNNs and Transformers to effectively capture low-level image features while modelling long-range cross-modal dependencies. The model was evaluated on two datasets: parotid gland tumour classification and knee injury classification. Parvaiz et al. (2023) surveyed the use of ViTs across various medical computer vision tasks, including image-based disease classification, anatomical structure segmentation, registration, region-specific lesion detection, captioning, report generation, and image reconstruction using multiple medical imaging modalities. These applications greatly support medical diagnosis and treatment. The authors also provide an overview of common imaging modalities used in medical computer vision and offer a concise explanation of the Transformer self-attention mechanism to enhance understanding. Tummala et al. (2022) proposed an ensemble of standard ViT models for the diagnosis of brain tumours. Zhang et al. (2025) proposed a transformer-based total knee replacement prediction using MRI. Arfat Ahmad Khan et al. (2024) proposed a mixed transformer-based model for semantic segmentation and triplet pre-processing of AD detection in a multi-class scenario. Rastogi et al. (2025) proposed a DL framework that integrates multiple MRI sequences to accurately segment brain tumours and predict survival outcomes in glioma patients. Aung et al. (2025) an enhanced U-Net hybrid model incorporating multiple attention mechanisms to strengthen feature representation and improve the accuracy of lung nodule segmentation. Kujur et al. (2022) presented a work on evaluating performance as a function of data complexity in brain MRI–based classification of brain tumors and AD. Unlike CNNs, which rely on local receptive fields and hierarchical feature extraction, ViTs use self-attention mechanisms that capture long-range dependencies across an entire image, enhancing performance in tasks that require understanding of relationships between distant parts of an image (Bazi et al., 2021).

ViTs scale more efficiently with increased data and computational resources, benefiting significantly from large-scale pre-training on massive datasets, unlike CNNs, which often require extensive architecture tuning and careful resource management to scale effectively. This work presents a discrete wavelet transform (DWT) and ViT-based model for AD detection. The remainder of the paper is organised as follows: The description of the database is in Section 2 followed by the methodology in Section 3. The results are presented in Section 4 followed by a discussion in Section 5. Finally, the paper is concluded in Section 6.

2 Database description

This study, which dealt with the identification of AD used data from the Alzheimer’s Disease Neuroimaging Initiative (ADNI), a database that can be found at (adni.loni.ucla.edu). The Food and Drug Administration (FDA), the National Institute of Biomedical Imaging and Bioengineering (NIBIB), private non-profit organisations, and commercial pharmaceutical corporations introduced the ADNI dataset in 2003. The primary function of the ADNI archive has been the analysis of its images, including serial MRI, PET, and various clinical, neuropsychological, and biological markers. In this work, intensity-normalised grayscale MRI and PET slices from this database have been used for the detection of AD. 77 AD, 80 MCI, and 82 control neutral (CN) image slices were used from phases (ADNI-1, GO, 2).

3 Proposed methodology

This section describes the adopted methodology used for AD detection. The proposed methodology is represented graphically in Figure 1 and the architecture of ViT is shown in Figure 2.

3.1 Image fusion

Radiological images play a crucial role in medical diagnosis by providing visual interpretation and evaluation. Integrating information from many modalities can help physicians make more informed decisions about therapy and surgery (Wald, 1999). MRI provides valuable information on soft tissue areas, including normal and diseased tissues. PET images show organ metabolic processes, such as blood flow and food activity, with limited spatial resolution (Leahy and Yan, 1991). The overarching image fusion scheme of the proposed algorithm comprises three key steps: initially decomposing the input registered images, followed by the application of fusion rules, and finally reconstructing the resultant composite fused image, as depicted in Figure 3.

The DWT offers significant advantages in the context of MRI and PET image fusion for AD detection. By enabling multiresolution analysis, DWT effectively decomposes images into various frequency components, capturing both high-frequency details (such as edges) and low-frequency information (such as smooth regions) (Subramanyam Rallabandi and Seetharaman, 2023). This capability is crucial for merging the anatomical details from MRI with the functional information from PET scans. The localised frequency information provided by DWT ensures that spatial variations are well-handled, preserving essential features from both imaging modalities. Additionally, DWT’s sensitivity to different directions helps in better representing intricate details, while its noise reduction capability enhances the clarity of the fused image by separating noise from useful signals. The computational efficiency of DWT allows for fast processing, which is beneficial in clinical settings. Importantly, DWT preserves edges and significant features, maintaining the diagnostic value of the fused image (Subramanyam Rallabandi and Seetharaman, 2023). Its robustness to small misregistrations between MRI and PET images further enhances its applicability. The ability to compact most of the image’s energy into a few significant coefficients ensures that the most important features are emphasised, leading to improved detection and analysis of AD.

During the implementation stage, it is crucial to handle the approximation C A i or C B i and detail D A i or D B i images differently due to their distinct physical significance. Detail images contain perceptually relevant information primarily related to edge details present in their coefficients. In our study, DWT and inverse discrete wavelet transform (IDWT) with Daubechies 2 (db2) wavelet was used for image fusion. Coefficients with large absolute values indicate abrupt intensity shifts, suggesting conspicuous features such as edges, lines, and area borders within the picture (Amolins et al., 2007). However, the nature of the approximation coefficients varies. These are a crude representation of the original picture, which may maintain attributes like mean intensity or coarse texture information. Using this information, the proposed fusion rules can be delineated as follows (Equations 1, 2) (Das and Bhattacharya, 2015):

C O i x , y = A v e r a g e C A i x , y , C B i x , y (1)

D O i x , y = M a x D A i x , y , D B i x , y (2)

Here, superscript i denotes the i t h level of resolution. To analyse the correlation between the original MRI/ PET image and the DWT-fused image two-dimensional correlation coefficient between the two images was calculated. It measures how strongly the pixel intensities in one image vary in relation to the pixel intensities in the other. Where a value close to 1 indicates a strong positive correlation (high similarity), a value close to 0 indicates no linear relationship, and a value close to −1 indicates a strong negative correlation (Equation 3).where A ̄ and B ̄ represent the mean intensities of images A and B, respectively. For this study, A was MRI/ PET image and B was the DWT fused image to calculate the correlation between MRI/PET and the Fused image. For the images shown in Figure 3, the correlation between MRI and the Fused image was 0.8801, and between PET and the fused image, it was 0.9404. Thus, these values show that the DWT fusion process incorporates both MRI and PET features.

corr 2 A , B = ∑ i = 1 m ∑ j = 1 n A i j − A ̄ B i j − B ̄ ∑ i = 1 m ∑ j = 1 n A i j − A ̄ 2 ∑ i = 1 m ∑ j = 1 n B i j − B ̄ 2 (3)

3.2 Image augmentation

This section describes the pre-processing performed on the image before it was applied to the ViT network. We created an ‘augmented image database’ from these resized images. The image augmentation techniques used in this work are rotation, reflection, translation, scaling, and shear. The original image and the augmented images are shown in Figure 4. Each image resulted in five augmented images after this step. This gives rise to 462 AD images and 492 healthy images.

3.3 Vision information transformer (ViT)

The ViT is a huge step forward in computer vision, employing the transformer architecture initially designed for natural language processing applications. Traditional CNNs have long been the main technique for image recognition tasks, but ViT offers an alternative paradigm that has received significant attention and success. The key idea behind ViT is to treat an image as a sequence of tokens, akin to the words in a sentence, and process it using the transformer architecture. ViT work flow is as flows (Dosovitskiy et al., 2020):

• Patch Embedding: Patch embedding is a crucial step in the ViT architecture, where an input image is transformed into a sequence of fixed-size patches, each of which is then embedded into a vector representation. This process enables the transformer to process the image as a sequence, similar to how it processes sequences of words in natural language processing tasks. The detailed breakdown of the patch embedding process is as follows.
• Image to Patches:
• An input image I of size H × W × C is divided into a grid of non-overlapping patches.
• Let P × P is size of each patch. Then the number of patch N = H P × W P.
• Flattening Patches: Each patch is flattened into a vector. If the patch size is P × P and the number of channels is C, each patch is transformed into a vector of size P 2 × C.
• Linear Embedding: Each flattened patch is then linearly embedded into a D-dimensional vector. This is done using a trainable linear projection E.
• Adding Positional Embeddings: To retain spatial information, positional embeddings are added to each patch embedding.
• Creating the Sequence: The sequence of embedded patches forms the input to the transformer encoder.
• Transformer Encoder: The Transformer Encoder is a critical component of the ViT, based on the transformer architecture initially created for natural language processing. It is made up of a stack of identical layers, each with two primary sub-layers: Multi-Head Self-Attention and a Feed-Forward Neural Network. Layer normalisation and residual connections are also applied after each sub-layer. The steps involved at transformer encoder are as follows:
• Input Sequence: The input to the transformer encoder is a sequence of embedded patches and, optionally, a classification token.
• Multi-Head Self-Attention: The first sub-layer in each encoder layer is the multi-head self-attention mechanism. It allows the model to focus on different parts of the sequence simultaneously.
• Feed-Forward Neural Network (FFN): The second sub-layer is a position-wise feed-forward network applied independently to each position:
• Stacking Encoder Layers: The process described above is repeated for each of the layers in the transformer encoder.
• Classification Head: The classification head in the ViT is responsible for producing the final classification output from the sequence of encoded patches and the classification token. This head is typically a simple fully connected (dense) layer applied to the representation of the classification token. Here’s a detailed breakdown of the classification head in the ViT.
• Input to the Classification Head: After passing through the transformer encoder layers, the input sequence contains the representations of all the patches and the classification token.
• Extracting the Classification Token: The representation of the classification token, denoted as CLS, is extracted at this stage.
• Fully Connected Layer: A fully connected layer (also known as a dense layer) is applied to produce the final classification logits.
• Softmax Activation: A softmax activation function can be applied to the logits to obtain the probabilities of each class.

ViT offers several advantages over traditional CNN-based approaches (Maurício et al., 2023; Han et al., 2022):

• Flexible Input Sizes: Unlike CNNs, which typically require fixed-size input images, ViT can process images of arbitrary sizes by dividing them into patches. This flexibility makes ViT more adaptable to a wide range of input resolutions and aspect ratios.
• Global Context Awareness: The self-attention mechanism in the transformer architecture enables ViT to capture global dependencies across the entire image, allowing it to leverage long-range spatial relationships effectively. This global context awareness is particularly beneficial for tasks requiring holistic understanding of the image.
• Parameter Efficiency: ViT tends to be more parameter-efficient compared to CNNs, especially for large input images, as it processes the image in a tokenized form rather than relying on convolutional filters. This can lead to significant computational savings, especially in scenarios with limited computational resources.
• Interpretability: The attention mechanism in ViT provides interpretability by allowing visualization of the attention weights, which indicate which parts of the image are relevant for making predictions. This transparency can be valuable for understanding the model’s decision-making process and diagnosing potential issues.

Since its introduction, ViT has achieved state-of-the-art performance on various image classification benchmarks, rivaling or surpassing the performance of CNN-based models in many cases.

4 Results

A variety of measures may be used to evaluate categorization performance. In this study, the following metrics are widely used: Sensitivity (Se), specificity (Sp), positive predictive rate (PPR), accuracy (Acc), Negative Predictive Value (NPV), False Positive Value (FPV), False Negative Nalue (FNV), F1 Score (F1), and Youden’s J statistic (J-stat). These together make up the system’s performance metrics. Mathematically, Sp and Se assess a test’s capacity to identify the presence or absence of an event. In this study, AD detection is classified “positive,” whereas normal circumstances are labelled “negative.” PPR is the proportion of true positives among positive results of diagnostic tests and statistics. The J-Stat statistic summarises the performance of a diagnostic test. All performance measures are evaluated as follows (Equations 4–12):

Sensitivity Se = T P T P + F N × 100 (4)

Specificity Sp = T N T N + F P × 100 (5)

Positive predictive rate PPR = T P T P + F P × 100 (6)

Accuracy Acc = T P + T N T P + T N + F P + F N × 100 (7)

Negative Predictive Value NPV = T N T N + F N × 100 (8)

False Positive Value FPV = F P T N + F P × 100 (9)

False Negative Nalue FNV = F N T P + F N × 100 (10)

F 1 Score F 1 = 2 × T P 2 × T P + F P + F N × 100 (11)

Youden ′ s J statistic J - stat = S e + S p − 1 × 100 (12)

Here,

• True positive( T P ): Number of positive predictions made when positive labels are actually present.
• False positive( F P ): Number of positive predictions made when negative labels are actually present.
• True negative( T N ): Number of negative predictions made when negative labels are actually present.
• False negative( F N ): Number of negative predictions made when positive labels are actually present.

After preprocessing, all images were divided into training and test sets with an 80:20 ratio. Then, DL models are trained using train set, using hyper-parameters defined in the Table 1. After training, models are evaluated against a test set of images. This experiment is performed 5 different time and results during these trails are are presented in Table 2.

TABLE 1. Hyper-parameters used in the proposed ViT.

TABLE 2. Experimental results for AD detection using ViT and other state-of-art Models.

Bold values indicates average of five experiments.

A comprehensive report is presented in Table 2. In addition to ViT, the fused images were also trained and tested on bench-marked transfer learning models: AlexNet, ResNet50, VGG19, ResNet101. The proposed ViT outperforms the other architectures in terms of all performance measures. ViT achieved an average Acc = 97.68%, PPR = 97.82%, Se = 97.39%, F1 = 97.75%, Sp = 97.96%, FPR = 2.04%, FNR = 2.61%, J-stat = 95.35%. AlexNet, ResNet50, VGG19 and ResNet101 achieved an average accuracy of 92.21%, 92.83%, 93.89%, and 95.05%. Performance of ViT in detecting AD over all the given performance measures is outstanding compared to the other models tested on the same dataset. This shows its efficacy and reliability in AD detection.

The receiver operating characteristic curve (ROC curve) is a graph that depicts how well a classification model performs at different levels of categorization. This curve represents two parameters: FPR and TPR. When the classification threshold is reduced, more objects are classified as positive, increasing the frequency of false positives as well as real positives. Figure 5 depicts the ROC curve for the ViT model corresponding to Experiment-1 in Table 2. ViT achieved an Area Under the Curve (AUC) of 0.99, which implies that this model is extremely effective at distinguishing between patients with Alzheimer’s and those without it. This high AUC suggests that the model will produce very few false positives and false negatives, leading to more accurate and reliable diagnoses.

Training and validation curves are graphical representations of the performance of a machine learning model during the training process. They help diagnose various aspects of the model. The training and validation corresponding to Experiment-1 in Table 2 is shown in Figure 6.

As per the statistical analysis done in Table 3, ViT clearly has the highest accuracy and narrow CI. AlexNet has the lowest mean accuracy ResNet101 and VGG19 perform strongly, but below ViT. Also, to test whether the mean accuracies differ significantly across models, a one-way ANOVA was performed. Since p < 0.00000001, the differences in accuracy between the models are statistically significant. This clearly shows that ViT performed better compared to the other models.

TABLE 3. Statistical summary of classification accuracy across models.

5 Discussions

5.1 Model parameter tuning

ViT performance can be significantly influenced by both batch size and patch size. These hyper-parameters play crucial roles in determining the generalizability of the model. The small batch size offers a noisy estimate of the gradient, which can help escape local minima and potentially lead to better generalization. However, it may result in slower convergence and less stable training. However, a large batch size provides a more accurate estimate of the gradient, leading to faster and more stable convergence. However, it can also lead to poorer generalization and the need for more sophisticated learning rate schedules. The proposed model is tested batch size of 8, 16, 32, and 64. The best performance was achieved at batch size of 32 which is represented in Figure 7.

Similarly, small patch size leads to a larger number of patches, which can provide more detailed local information and potentially better feature representation. However, it increases the computational load and memory usage due to the increased sequence length. However, large patch size reduces the number of patches, which decreases the computational load and memory usage. However, it may miss fine-grained details in the image, potentially reducing the model’s ability to capture intricate patterns. An analysis of different patch size is shown in Figure 8. Patch size of 7 × 7 is best for the proposed model for detection of AD.

5.2 Ablation study

An ablation study in deep learning is a systematic way to understand the contribution of various components of a neural network model to its overall performance. It involves selectively removing or modifying parts of the model and observing the resulting performance changes. This helps identify which components are crucial and how different elements of the model interact. To prove the effectiveness of DWT based image fusion in AD detection, the experiments were also done using a uni-model imaging system (MRI or PER only). Table 4 shows that the fusion of MRI and PET resulted in higher performance.

TABLE 4. Ablation study of the proposed model for DWT based image fusion and ViT based classification.

5.3 Comparison with existing state-of-art techniques

This research describes a ViT-based model for diagnosing Alzheimer’s disease. In recent years, several ML or DL-based approaches for AD detection have been presented. This section contains a thorough explanation and comparison of the planned work to existing state-of-the-art models.

Akhila et al. (2017) introduced a neural network-based approach. This paper describes a method for properly detecting dementia using brain MRI. Feature extraction is performed using the segmentation-based fractal texture analysis (SFTA) approach. The binary images generated by the Two Threshold Binary Decomposition approach, once the picture has been broken down, are used to extract fractal dimensions and texture attributes. Dementia is classified using neural networks. This approach was evaluated on 3D brain MRI images from the OASIS dataset and achieved a classification accuracy of 97.5%. Faturrahman et al. (2017) proposed a Deep Belief Network (DBN)-based categorisation technique for Alzheimer’s disease that relies on structural modalities and offers strong predictive models. Support vector machines (SVM) have been compared to DBN in classification models. The results show that this technique outperforms SVM and the current method in the previous examination. The DBN has Acc, Se, and Sp values of 91.76%, 90.59%, and 92.96%, respectively. To reduce the dimensionality of retrieved features from MRI neuroimages, E.M. (Alkabawi et al., 2017a; Alkabawi et al., 2017b) proposes a unique feature abstraction approach employing sparse autoencoder (SAE). Then, to evaluate the prediction accuracy of a linear discriminant analysis (LDA) and a logistic regression classifier, these features, as well as those obtained by the well-known PCA technique, are used to train the corresponding classifiers. The experimental findings show that the suggested technique improves classification accuracy by 8% compared to the PCA. Cui et al. (2018) propose a classification framework for long-term assessment of structural MR images in AD diagnosis, using a mix of convolutional and recurrent neural networks. CNNs are first created to learn the spatial features of MR images for classification purposes. To extract longitudinal characteristics for AD classification, recurrent neural networks (RNNs) with cascading three bidirectional gated recurrent units (BGRU) layers are created on CNN outputs at different time periods. Experiments show that the recommended approach achieves a classification accuracy of 91.33%. To predict the transition from MCI to AD, a data mining framework was designed and combined with three different classifiers: SVM, Bayes statistics, and voting feature intervals (VFI) (Plant et al., 2010). Using feature selection approaches, the brain areas with the best accuracy for discriminating between AD and CN were discovered, with up to 92% classification accuracy. Liu et al. (2013) propose a novel diagnostic technique called Multifold Bayesian Kernelization (MBK), which represents the diagnosis process as a synthesis of multi-modal biomarkers. Each biomarker is assigned an MBK kernel that maximises local neighbourhood affinity, and the contribution of each biomarker is then evaluated using a Bayesian framework. By integrating the output diagnostic probabilities of numerous biomarkers, MBK produces a unique diagnosis scheme that may help identify the subject’s condition. Papakostas et al. (2015) suggested a strategy based on Lattice Computing (LC) scheme for detecting AD using MRI data. A k-NN classifier in the LC setting approaches computer-assisted diagnostics from two different perspectives. To categorise the subjects inside the lattice space, it first performs dimensionality reduction on the high dimensional feature vectors. H.I. Suk et al. (2017) propose a novel framework for diagnosing and predicting AD and MCI, using Deep Ensemble Sparse Regression Network (DESRN). To be more specific, train multiple sparse regression models first, each with a distinct value for the regularisation control parameter. C.V. Dolph et al. (2017) present a multiclass deep learning classification of Alzheimer’s disease using unique texture and associated factors from structural MRI. Afzal et al. (2019) employed a transfer learning-based strategy to enrich 3D MRI images from the OASIS dataset. The recommended model achieves 98.41% accuracy with a single brain MRI scan and 95.11% with 3D views. From structural MRI to diffusion tensor imaging, Karim et al. (2018) propose a cross-modal transfer learning approach. To train on mean diffusivity data, models trained on a structural MRI dataset with domain-dependent data augmentation are used to initialise the network parameters. The strategy eliminates overfitting, increases learning efficiency, and hence improves prediction accuracy. Glozman et al. (2017) developed a technique that uses statistical learning and projection-based shape descriptors to create a canonical representation of brain regions. This technique automatically identifies the structures that have been most badly harmed by the illness. Cheng et al. (2017) offer a novel multi-domain transfer learning (MDTL) architecture for learning tasks in several auxiliary domains as well as the target domain. Maqsood et al. (2019) suggested utilising 3D MRI scans to classify and identify AD phases using transfer learning. The capacity to identify AD through the integration of multimodal neuroimaging data has been established utilising methods like PET and MRI using Multi-modal stacked deep polynomial networks (MM-DPN) (Shi et al., 2017). Payan and Montana suggested a deep 3D CNN that can learn generic features collecting AD biomarkers and discriminate between Alzheimer’s brain and a normal healthy brain based on brain MRI data to predict AD. Ortiz et al. (2016) investigate how deep learning architectures may be utilised to create classification algorithms for areas of the brain identified by Automated Anatomical Labelling (AAL). The AAL atlas regions have been used to partition grey matter (GM) images from each brain region into 3D patches, which are then utilised to train a variety of deep belief networks. A complete deep learning system based on MRI and PET scans might be used to predict the course of AD. Li et al. (2015) improved the DL framework by introducing a Stacked Robust Deep Model (SRDM), an adjustable learning factor, and stability selection. Lu et al. (2018) propose a new deep-learning architecture that employs a multimodal and multiscale deep neural network to differentiate between persons with AD. This technique detects persons with MCI with an accuracy rating of 82.4%. As mentioned in this section and illustrated in Table 5, the suggested model outperforms the existing state-of-the-art methods based on ML or DL. When compared to ML-based models (Akhila et al., 2017; Horn et al., 2009; Gunawardena et al., 2017; Liu et al., 2013; Papakostas et al., 2015),The recommended technique performed better. DL models offer an advantage over ML models since they do not require feature engineering. Comparison with DL based models (Faturrahman et al., 2017; Alkabawi et al., 2017a; Alkabawi et al., 2017b; Cui et al., 2018; Gunawardena et al., 2017; Wang et al., 2018a; Dolph et al., 2017; Payan and Montana, 2015; Wang et al., 2018b; Lu et al., 2018) this method has significantly performed better. The suggested strategy is also compared with models that use DL in the feature extraction stage and ML in the classification step (Horn et al., 2009; Plant et al., 2010; Papakostas et al., 2015; Afzal et al., 2019; Ortiz et al., 2016). The suggested model outperformed these models. The suggested study employs a TL-based technique to AD detection. Compared with TL techniques (Afzal et al., 2019; Karim et al., 2018; Wang et al., 2017; Glozman et al., 2017; Dyrba et al., 2013; Maqsood et al., 2019; Cheng et al., 2017), This strategy has done exceptionally well. Xiaoke et al. (2020) has proposed a methods multi-modal neuroimaging feature selection method with consistent metric constraint for AD detection. The first step in this method for detecting AD using MRI and PET image fusion is to use a random forest model to calculate pairwise similarities for each modality separately. By efficiently extracting similarity metrics, this technique captures the connections between various types of data. Subsequently, the objective function is constrained using a sample similarity constraint regularisation term and a group sparsity regularisation term to carry out feature selection. These limitations aid in the process of choosing the most pertinent qualities among the many modalities. In order to achieve a final classification that is reliable and accurate, a multi-kernel support vector machine (MK-SVM) is used to fuse the selected features from various models. Odusami et al. (2023) suggested a DL-based system that combines PET and MRI data. In this work, PET and MRI images are concatenated using a heuristic early feature fusion methodology. A modified ResNet18 deep learning architecture is trained concurrently. In order to facilitate binary categorisation of AD, a 3-in-channel technique is applied to extract the most descriptive features from the merged PET and MRI images. Subramanyam Rallabandi and Seetharaman (2023) first perform 3D tissue segmentation on MR images after skull stripping. Following segmentation, PET images and atlas-based MR tissue images are combined. PET pictures are then converted from RGB to HSI colour space, and two-dimensional Fourier transforms and DWT are used to fuse MRI and PET images. After that, the MR-PET fused picture is rebuilt using DWT and inverse Fourier techniques. CNN were followed by this step for classification. Qi et al. (2025) proposed a hybrid technique for AD detection using a CNN-Transformer combination. In this architecture, the CNN is first used as a feature extractor to generate feature maps. This design choice has two key advantages: it enables the use of high-level features during decoding, and it allows the Transformer to better capture the spatial context of the patch features, leading to a more refined interpretation of the image. Lu et al. (2023) proposed AD detection using ConvNeXt and Dynamic Multilayer Perceptron (DMLP). Image features are extracted using the convolutional blocks and global average pooling layers of the ConvNeXt network. The subjects’ age and gender information is then parameterised using a DMLP, and these parameters are combined with the extracted features before being passed to a fully connected layer. Finally, the softmax layer produces the classification outputs. Compared with the other methods based on image fusion (Odusami et al., 2023; Subramanyam Rallabandi and Seetharaman, 2023; Xiaoke et al., 2020; Forouzannezhad et al., 2018), the proposed DWT-based image fusion and ViT-based classification have performed well. Improved precision in this work may help identify AD with greater confidence.

TABLE 5. Comparison of the proposed model with existing methods for AD detection.

5.4 Performance under different classification scenarios

Classification between MCI, AD, and CN is vital for early detection, effective treatment, and better understanding of disease progression. MCI often serves as a transitional stage between normal ageing and AD, making its accurate identification critical for predicting and potentially delaying the progression to Alzheimer’s. Early diagnosis provides a crucial window for intervention, allowing for the implementation of treatments and lifestyle changes that can slow cognitive decline. Differentiating these conditions ensures that people receive appropriate care tailored to their specific needs, whether it involves preventive measures for those with MCI, targeted treatments for AD patients, or reassurance for healthy individuals. In addition, accurate classification improves research on the underlying mechanisms of cognitive decline and improves the allocation of healthcare resources by focusing interventions where they are most needed. The model proposed in this paper has been tested for multi-class classification between MCI vs. AD vs. CN. For this three-class classification, the obtained results are presented in Table 6. Average Acc, Se, Sp of 76.57%, 76.6%, 90.0% has been achieved. ViTs are powerful for capturing global spatial relationships, but they still require large, diverse datasets to learn fine-grained distinctions. The sharp drop in performance from the binary task (CN vs. AD) to the three-class task (CN vs. MCI vs. AD) indicates that while the model can easily separate the two extremes of the disease spectrum, it struggles to detect the nuanced and often diffuse structural changes characteristic of MCI. This is clinically expected: MCI features are less pronounced, can overlap substantially with both CN ageing patterns and early AD pathology, and are known to introduce higher inter-subject variability.

TABLE 6. Performance analysis of the proposed model for multi-class classification (MCI vs. AD vs. CN).

The Table 7 presents the performance analysis of the proposed model in various classification scenarios to distinguish the stages of AD. Four scenarios are evaluated: binary classifications of CN vs. MCI, MCI vs. AD, CN vs. AD, and a multi-class classification involving CN, MCI, and AD. In the CN vs. MCI scenario, the model achieves an accuracy of 86.42%, with sensitivity and specificity values of 85.37% and 87.50%, respectively, indicating balanced performance in distinguishing normal individuals from those with mild cognitive impairment. For MCI versus AD classification, the model provides a comparable precision of 86. 67%, with slightly higher sensitivity (87.50%) and slightly lower specificity (85.71%), showing consistent performance in differentiating between MCI and AD. The CN vs. AD classification achieves the highest performance, with an impressive accuracy of 97.68%, sensitivity of 97.39%, and specificity of 97.96%. This indicates the model’s exceptional ability to differentiate between cognitively normal individuals and those with AD. However, in the multiclass classification scenario involving CN, MCI, and AD, the model performance drops to an accuracy of 76. 57% and a sensitivity of 76. 60%, while maintaining high specificity at 90.00%. This suggests the added complexity of multi-class classification poses challenges to the model, leading to a reduction in overall accuracy and sensitivity, though the specificity remains strong.

TABLE 7. Performance analysis of the proposed model under different classification scenarios.

Overall, the results highlight the model’s strength in binary classification, especially in distinguishing CN from AD. A comparison with the existing fusion based methods for this multi-class classification is presented in Table 8. Song et al. (2021) have proposed a 3D multi-scale CNN to evaluate the performance of their image fusion algorithm in binary classification and multi-class classification tasks. They have achieved 3-class classification accuracy of 74.54%. Liu et al. (2014) have proposed a multimodal neuroimaging feature learning for multiclass detection of AD and achieved classification accuracy of 53.79%. Our image fusion algorithm demonstrated considerable benefits in the multi-class classification task.

TABLE 8. Comparative analysis of the proposed model with existing fusion-based methods for multi-class classification (MCI vs. AD vs. CN).

5.5 Ablation study

To understand the contribution of DWT-based image fusion in accurate classification of AD an ablation study was performed. In this experiment, MRI and PET images were fused using pixel averaging rather than the proposed DWT-based fusion. The remaining structure of the proposed model was kept as it is; the results obtained for AD detection are presented in Table 9. The drop in Acc when using pixel averaging clearly shows the potential of DWT-based fusion of MRI and PET images. Averaging combines MRI and PET by taking the pixel-wise mean, which often suppresses important structural or metabolic details because high-intensity regions in one modality may be diluted by low intensities in the other. In contrast, DWT fusion decomposes each modality into multi-resolution sub-bands. DWT-based fusion produces a more discriminative representation than simple averaging, enabling higher accuracy and more reliable AD detection.

TABLE 9. Ablation study: Result with and without DWT-based image fusion.

5.6 Limitations and future scope

• The proposed pipeline—combining DWT-based image fusion, data augmentation, and training a ViT introduces substantial computational overhead. ViTs in particular are known for their high memory consumption and long training time.
• This model is implemented on computer-aided system with images from a standard dataset. Implementing in a real-time system with clinical validation will be an interesting project.
• In this work, augmentation is done before split, which may lead to potential information leakage.
• The proposed model was implemented on a limited dataset. A wider analysis on a larger and more balanced dataset can bring more generalisation to the results and model parameters.
• DB2 wavelets were used because of their simple structure and ability to provide excellent edge preservation and noise reduction. They capture fine anatomical details while maintaining computational efficiency. Other wavelet families, such as Symlets, and Coiflets, can also be explored for MRI and PET analysis.

6 Conclusion

Alzheimer’s disease (AD) is the most prevalent form of dementia, distinct from natural aging, and a leading cause of death among seniors in many countries.This research aims to develop a highly accurate AD detection model using information from both magnetic resonance and PET images. In this study, PET and magnetic resonance images are fused using the discrete wavelet transform. Following data augmentation, these fused images are fed into a ViT based model for classification. A detailed analysis of model tuning of ViT, i.e. batch size and patch size is presented. Through detailed analysis, it was found that batch size = 32 and patch size = 7 × 7 gives the best performance. A comprehensive analysis is presented using 9 different performance measures: Se, Sp, PPR, Acc, NPV, FNV, F1, AUC, and J-stat. The ViT model outperformed the state-of-the-art techniques, by achieving an accuracy of Se = 97.39%, Sp = 97.96%, PPR = 97.82%, Acc = 97.68%, FPR = 2.04%, FNR = 2.61%, F1 = 97.75%, AUC = 0.99, and J-stat = 95.35%. The ablation study, multi-class classification and comparison with existing methods indicates that the DWT based fused information from MRI and PET images together with ViT can contribute significantly to AD detection.

Generative AI statement

The author(s) declared that generative AI was not used in the creation of this manuscript.

Any alternative text (alt text) provided alongside figures in this article has been generated by Frontiers with the support of artificial intelligence and reasonable efforts have been made to ensure accuracy, including review by the authors wherever possible. If you identify any issues, please contact us.

Publisher’s note

All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.

Footnotes

Edited by: Adriano De Oliveira Andrade, Federal University of Uberlândia, Brazil

Reviewed by: Arfat Ahmad Khan, Khon Kaen University, Thailand

Ketki Pathak, Sarvajanik College of Engineering and Technology, India