INTRODUCTION
Among the most common neurodegenerative diseases worldwide, Alzheimer’s disease (AD) causes a gradual loss of cognitive function and memory. With the aging of populations, the occurrence of AD is projected to increase dramatically, presenting substantial difficulties for healthcare systems around the globe (Grueso and Viejo-Sobera, 2021). The early identification of AD is essential for prompt intervention and treatment planning; however, it remains a challenging task owing to the subtle and diverse character of its beginning. Advancements in machine learning (ML) and deep learning (DL) have shown potential in enhancing the accuracy and efficiency of AD detection in recent years (Hedayati et al., 2021).
Timely identification enables prompt implementation of intervention techniques and treatment planning, mitigating illness development and enhancing patient outcomes (Zhao et al., 2023). The effectiveness of interventions such as lifestyle adjustments, medication management, and cognitive training is most when they are started early in the course of the illness. Timely diagnosis allows persons with AD and their families to proactively plan for the future, make well-informed choices about treatment alternatives, and avail themselves of support resources. It mitigates feelings of fear and confusion associated with cognitive loss by offering a more lucid comprehension of the underlying disease (Savaş, 2022). Precise and prompt diagnosis is crucial for enlisting volunteers in clinical trials for innovative treatments and medication development. Identifying people at an early stage who are at risk of developing AD may help in the evaluation of prospective therapies that are designed to prevent or postpone the start of the illness (Savaş, 2022). Earlier diagnosis decreases healthcare expenditures related to AD by mitigating the need for hospitalizations, visits to the emergency department, and long-term care costs (Balaji et al., 2023). Additionally, it enables the optimization of resource allocation and healthcare planning.
Magnetic resonance imaging (MRI) enables the observation of both structural and functional alterations in the brain that are linked to AD, typically occurring prior to the appearance of clinical symptoms (Koga et al., 2022). Automated systems have the capability to examine MRI scans in order to identify minor anomalies that are symptomatic of AD pathology in its early stages. This allows for prompt intervention and therapy. Conventional approaches to diagnosing AD depend on subjective assessment of clinical symptoms and neuropsychological testing, which are susceptible to inconsistencies and prejudice (Sekhar and Jagadev, 2023). Automated techniques provide an impartial and uniform method for detecting AD, decreasing the need for subjective assessment and improving the accuracy of diagnosis.
This study introduces an innovative approach for detecting AD using ensemble learning (EL). The method combines the advantages of ML and DL techniques (Liu et al., 2020). The integration of YOLOv7 and EfficientNet B3 is used to extract features, using CatBoost and XGBoost as base learners and support vector machines (SVMs) as meta-learners. The objective of our ensemble framework is to use the combined advantages of these techniques to improve the accuracy of AD detection while also ensuring resilience and generalization.
DL algorithms, such as YOLOv7 and EfficientNet B3, represent a notable shift away from conventional approaches to feature extraction in identifying AD (Loddo et al., 2022). The YOLOv7 algorithm allows for accurate identification and extraction of important characteristics from brain imaging data. EfficientNet B3 can overcome the challenges in image classification, capturing subtle patterns and employing semantic information to differentiate AD and healthy brain images (El-Sappagh et al., 2022). It can build hierarchical representations using raw data.
The ensemble framework incorporates CatBoost and XGBoost as base learners, in addition to DL approaches. Gradient-boosting algorithms have shown exceptional performance in many classification problems due to their capability to manage intricate relationships between features and reduce overfitting (Tanveer et al., 2020). Moreover, SVMs serve as meta-learners in order to combine the outputs of the base learners and enhance the decision boundaries for enhanced classification (Loddo et al., 2022). SVMs are particularly effective in dealing with data with many dimensions and complex, nonlinear connections. This makes them very suitable for combining predictions from other classifiers. By using SVMs as meta-learners, the objective is to use their capacity to maximize classification margins and enhance the overall generalization performance of the ensemble model.
The main goal of this project is to build a strong and precise system for detecting AD that can effectively use the advantages of both DL and ML approaches. The study objective is to obtain exceptional performance in distinguishing between AD and non-AD individuals by using YOLOv7 and EfficientNet B3 for feature extraction, along with CatBoost, XGBoost, and SVMs in an EL approach. In addition to improving classification accuracy, the suggested ensemble structure encourages model resilience and generalization, enabling more dependable and understandable diagnostic choices.
The critical contributions of the proposed study are as follows:
Integration of DL and ML techniques: This study proposes a novel EL approach that integrates DL techniques, including YOLOv7 and EfficientNet B3, with traditional ML algorithms, such as CatBoost, XGBoost, and SVMs. By combining the strengths of these diverse algorithms, we aim to enhance the accuracy and robustness of AD detection models.
Feature extraction using DL models: YOLOv7 and EfficientNet B3 are employed for feature extraction from brain imaging data, enabling the capture of spatial and semantic information crucial for AD detection. By leveraging the hierarchical representations learned by these DL models, we aim to improve the discriminatory power of the extracted features.
EL paradigm: CatBoost and XGBoost are utilized as base learners in conjunction with SVMs as meta-learners within an EL framework. This ensemble approach facilitates the combination of diverse predictions from multiple classifiers, leading to enhanced classification performance and improved model generalization.
Achievement of high accuracy: Experimental results on a Kaggle repository dataset demonstrate the efficacy of the proposed ensemble framework, with an average accuracy of 99.8% in AD detection. This remarkable performance underscores the potential of our approach in aiding clinicians with early diagnosis and intervention strategies for AD.
RELATED WORKS
The rising global incidence of AD makes it a significant public health concern. Timely intervention and treatment planning needs the early and precise diagnosis of AD (Nguyen et al., 2022). However, standard diagnostic approaches have limits in terms of their sensitivity, specificity, and scalability. Automated techniques for detecting AD utilizing MRI images have recently emerged and have shown the potential to enhance diagnostic precision and efficiency (Sadat et al., 2021). MRI provides valuable information on the structural and functional alterations in the brain that are linked to the pathology of AD. Several MRI biomarkers, including hippocampus volume, cortical thickness, and white matter integrity, have been recognized as adequate indications of disease development. Automated methods use these biomarkers to identify minor irregularities that indicate AD at its early stages before clinical symptoms become evident.
Wang et al. (2020) created an ML model that integrated many MRI indicators to effectively differentiate between patients with AD and healthy individuals, achieving a high level of sensitivity and specificity. DL methods, namely convolutional neural networks (CNNs), have become popular in the automated identification of AD using MRI scans. CNNs can autonomously acquire hierarchical representations from unprocessed image input, effectively capturing intricate patterns and characteristics that are pertinent to AD pathogenesis.
EL techniques, which merge predictions from several classifiers, have shown efficacy in enhancing the accuracy of AD detection. Sudharsan and Thailambal (2023) proposed a comprehensive framework that included many MRI biomarkers obtained via distinct feature selection techniques and classification algorithms. Their methodology resulted in superior classification accuracy and resilience compared to individual classifiers, emphasizing the advantages of EL in diagnosing AD. The integration of multimodal imaging data, including MRI, positron emission tomography (PET), and cerebrospinal fluid (CSF) biomarkers, has shown potential in improving the performance of AD diagnosis. Automated methods that integrate data from several modalities may provide a more thorough evaluation of AD pathology and enhance the precision of diagnosis. Ma et al. (2023) introduced a multimodal fusion framework that integrated MRI and PET imaging data using a DL-based fusion network. This technique demonstrated better performance in AD classification compared to utilizing just one kind of imaging data. Although automated AD detection methods have shown promising results in research settings, it is crucial to validate and translate them for real-world clinical use (Fang et al., 2020; AlSaeed and Omar, 2022). It is necessary to conduct extensive investigations involving several centers in order to assess how well these systems may be applied to different demographics and imaging platforms. Moreover, obtaining regulatory permission and successfully integrating it into clinical practice need thorough validation and compelling proof of therapeutic usefulness. There are still several obstacles to overcome in the creation and implementation of automated methods for detecting AD using MRI scans. These factors include the need for uniform imaging procedures, synchronization of data across several collection locations, comprehensibility of DL models, and ethical concerns involving data confidentiality and patient agreement. Potential areas for future study are the advancement of explainable artificial intelligence (AI) methods to improve the interpretability of models, the incorporation of longitudinal imaging data for monitoring the course of diseases, and the investigation of innovative imaging biomarkers for the early identification and prediction of AD.
MATERIALS AND METHODS
Figure 1 reveals the proposed AD detection model. The novelty of the proposed model lies in employing a hybrid feature extraction and EL approach for AD detection. Hybrid feature extraction algorithms, including YOLOv7 and EfficientNet B3, retrieve complementary information from MRI images. YOLOv7 collects spatial information about crucial anatomical structures in object identification and localization, whereas EfficientNet B3 captures semantic information and subtle AD disease patterns. Combining these two techniques allows the model to use spatial and semantic information to distinguish AD groups and phases. EL methods like CatBoost, XGBoost, and SVM models improve AD detection system resilience and generalization.
The proposed AD detection framework. Abbreviations: AD, Alzheimer’s disease; SVM, support vector machine.
The ensemble may reduce model biases and mistakes by aggregating their predictions, making them more accurate and dependable. Ensemble approaches can reduce overfitting by utilizing many models trained on distinct data subsets or methodologies. The ensemble is diverse since CatBoost, XGBoost, and SVM models use different ML techniques. SVMs excel at high-dimensional data and nonlinear correlations, whereas CatBoost and XGBoost are gradient-boosting algorithms that manage complicated feature interactions and reduce overfitting. The ensemble can capture more data patterns and correlations by mixing models with various properties, increasing performance. EL routinely outperforms solo models in ML problems. The ensemble may improve AD detection accuracy and reliability by integrating the predictions of many models trained on various feature representations or learning strategies. Minor performance gains might have major clinical ramifications in medical applications where diagnosis accuracy is crucial.
EL can enhance AD detection system’s interpretability and explainability. Researchers may learn about categorization judgments and which elements are most essential for identifying AD groups or stages by pooling predictions from numerous models. Transparency helps create confidence in automated diagnostic technologies and promotes clinical adoption.
The choice of CatBoost, XGBoost, and SVM as base learners and meta-learners in an ensemble model for AD detection from MRI data is justified by their individual strengths and complementary capabilities. CatBoost’s efficiency and categorical handling, XGBoost’s high performance and regularization, and SVM’s robustness in high-dimensional spaces and interpretability together create a powerful and well-rounded ensemble architecture. This combination ensures improved accuracy, generalizability, and robustness, which are critical for reliable AD detection in clinical practice.
Dataset acquisition
The open access series of imaging studies (OASIS) MRI dataset contains 80,000 brain MRI images. The images have been organized into four AD progression groups (Kaggle, n.d.). The dataset is intended for AD analysis and detection. The dataset owner used FSL (FMRIB Software Library) to convert the original “.img” and “.hdr” files into Nifti format (.nii), which made the dataset available. The transformed MRI images of 461 patients are available in several sections on GitHub. Neural networks were trained on two-dimensional pictures. Patients’ brain scans were cut along the z-axis into 256 parts, with 100-160 slices chosen. This method yielded a large dataset for study. The metadata and Clinical Dementia Rating values classified patients as demented, very mildly demented, mildly demented, or non-demented. These seminars identify and examine AD phases. The dataset preparation included converting “.nii” MRI images to “.jpg” files. Although this translation was difficult, the files were handled using proper tools. The authors used the large dataset to test neural network models for AD detection and analysis.
Feature extraction
The integration of YOLOv7 and EfficientNet B3 for feature extraction in MRI data analysis offers a powerful toolset for AD detection. YOLOv7’s strength in precise object detection and localization, combined with EfficientNet B3’s advanced feature extraction capabilities, results in a comprehensive, accurate, and efficient method for analyzing MRI data. This combined approach enhances the potential for early diagnosis, effective monitoring, and improved understanding of AD, ultimately contributing to better patient outcomes and advancing research in the field. Using YOLOv7 and EfficientNet B3 for feature extraction provides a robust and all-encompassing method for analyzing MRI data in the context of AD detection. By using spatial and semantic information, these designs extract meaningful characteristics that can assist in early detection, monitoring, and personalized therapy approaches for individuals with AD. YOLOv7 and EfficientNet B3 are specifically engineered to be scalable and highly efficient, making them well suited for effectively processing substantial amounts of MRI data. Their efficient designs allow for quick processing, making them suitable for real-time analysis and implementation in medical environments. Moreover, their effectiveness enables them to be easily adjusted to various hardware platforms and contexts with few resources, allowing widespread accessibility and application of the feature extraction pipeline for AD diagnosis utilizing MRI.
YOLOv7 model
YOLOv7 may be conveniently included in DL frameworks like TensorFlow or PyTorch, enabling quick training and deployment of object detection models for autonomous driving. YOLOv7 utilizes DL techniques to extract advanced features from MRI scans, revealing intricate patterns and correlations that may indicate AD disease. The authors load the weights of the pre-trained YOLOv7 model to initialize it specifically for feature extraction. The MRI scans were passed through the YOLOv7 model to extract features. During the forward pass, the YOLOv7 model utilizes a sequence of convolutional and pooling layers to process the input pictures, progressively extracting hierarchical features at various levels of abstraction. The authors applied Adam optimizer to fine-tune the hyperparameters of the YOLOv7 model to incorporate domain-specific knowledge to improve the accuracy and robustness of the feature extraction process.
EfficientNet B3 model
Researchers and practitioners may effectively utilize the EfficientNet B3 model to extract features from MRI scans. This can lead to the development of precise and dependable ways for detecting and diagnosing AD. The pre-trained EfficientNet B3 model weights were utilized to initialize it specifically for feature extraction. During the forward pass, the EfficientNet B3 model utilizes a sequence of convolutional and pooling layers to process the input pictures, extracting hierarchical features at various levels of abstraction. The model often produces feature maps comprising sophisticated input picture representations.
Within the framework of AD, precise identification and diagnosis heavily rely on extracting numerous features from MRI images. These aspects include alterations in the structure and function of the brain, which may serve as indicators of the course of the illness. A key characteristic is cortical atrophy in the hippocampus, entorhinal cortex, and other areas linked to memory and cognitive functioning. The hippocampus, a vital component in the process of creating new memories, is often one of the first regions to exhibit indications of degeneration in individuals with AD. Ventricular enlargement is a notable characteristic in which the fluid-filled chambers in the brain, known as ventricles, expand due to brain tissue loss. This phenomenon is observed in conjunction with cortical atrophy. White matter lesions, characterized by brain white matter wounds or degeneration, are frequently associated with AD. These abnormalities may interfere with the transmission of information across various parts of the brain, leading to a loss of cognitive function. Moreover, alterations in the thickness of the cortex, namely cortical thinning, in different areas of the brain serve as indicators of neuronal death and are used as markers for AD.
Feature fusion
Feature fusion integrates features obtained from several sources to generate a more comprehensive and valuable representation of the input data. In the context of AD detection, feature fusion combining EfficientNet B3 and YOLOv7 features can increase classification accuracy by using spatial and semantic information. The authors combined the feature vectors obtained by EfficientNet B3 and YOLOv7 by concatenating them along the feature dimension. This integrates both spatial and semantic information into a unified feature representation. Element-wise addition was utilized on the concatenated feature vectors to decrease redundancy and enhance computational efficiency. Figure 2 highlights the recommended feature fusion technique.
Ensemble learning-based AD classification
CatBoost model
CatBoost can be straightforwardly included in EL frameworks, functioning as a robust base learner that enhances the ensemble’s overall prediction accuracy. Ensemble models can achieve higher performance in AD detection by using the capabilities of individual models through the combination of CatBoost with other base learners and a meta-learner. The authors instantiated and trained a CatBoost classifier using the training data. Furthermore, they can establish hyperparameters such as the number of trees, learning rate, tree depth, and regularization parameters.
XGBoost model
The ultimate prediction in XGBoost is achieved by combining the predictions of multiple decision trees. Each decision tree is trained separately, but it contributes to the final prediction by giving a vote with a certain weight, determined by its performance on the training data. XGBoost offers a technique to assess the significance of each feature in the dataset. XGBoost computes the impact of each feature on the decrease in the loss function, such as the mean squared error, during the training phase. These data may be utilized to determine the most significant characteristics for detecting AD, offering a vital understanding of the fundamental patterns in the data. During the prediction procedure, every sample in the testing dataset is sequentially processed by the ensemble of decision trees. XGBoost computes the expected result for each sample by combining the predictions from all the decision trees, usually by a weighted average. The weights allocated to each tree are chosen by evaluating their performance on the training data and their contribution to minimizing the loss function.
SVM model
SVM identifies the effective hyperplane that can accurately distinguish distinct classes in the feature space. This hyperplane is chosen to maximize the margin, representing the distance between the hyperplane and the nearest data points, also known as support vectors. SVMs achieve a solid decision boundary that effectively generalizes to new data by maximizing the margin. SVMs commonly employ the one-versus-all (OvA) technique in multi-class classification applications. This strategy involves training individual binary classifiers for each class. Each classifier is responsible for distinguishing between one specific class and all the other classes. During the prediction phase, the trained SVM classifier calculates the decision function of each class by using the support vectors and their associated weights that were produced during the training procedure. When a novel point exists, the decision function calculates values for each class, and the class with the most significant value is chosen as the predicted class. The OvA technique involves the use of binary classifiers, each of which generates a decision function value. The predicted class is determined by selecting the class with the highest decision function value.
Performance validation
The validation set was used to assess the model’s performance, employing metrics including accuracy, precision, recall, specificity, intersection over union (IoU), and F1-score. Accuracy is the proportion of correctly classified instances out of the total number of cases. Precision is the ratio of true-positive predictions to the total number of positive predictions. It measures the model’s ability to identify positive instances while minimizing false positives correctly. Recall is the ratio of true-positive predictions to the total number of actual positive instances. It measures the model’s ability to capture all positive instances, minimizing false negatives. F1-score is the harmonic mean of precision and recall, providing a balanced measure of a model’s performance. It considers both false positives and false negatives and is especially useful when dealing with imbalanced datasets. IoU measures the overlap between the predicted and ground truth bounding boxes. IoU is calculated as the overlap area divided by the union area between the two boxes. A higher IoU indicates better localization accuracy.
RESULTS AND DISCUSSION
The authors present the experimental settings, results, and discussion in this section. They employed Ubuntu 23.10.0 (Mantic Minotaur) long term support, Intel (R) Core i5-7500 CPU @ 3.40 GHz, 16 GB DDR4 RAM, and Anaconda Python Environment to implement the proposed model. The dataset was organized into a train set (60%), a validation set (20%), and a test set (20%). A total of 79 epochs and 92 batches were used to train the model. The model has achieved an average accuracy of 99.7% on the validation set at the 69th epoch and 84th batch. However, the training was extended to 79 epochs and 92 batches in order to improve the model’s efficiency.
The findings of the performance validation are highlighted in Table 1. The presented model attained an average accuracy of 99.8% in multi-class classification, which is an exceptional result. This indicates that the suggested model for AD detection is resilient and effective. In order to confirm the accuracy of the model, we performed a performance assessment on the OASIS dataset. This dataset consists of many classes that reflect different levels of dementia severity, including mild dementia, moderate dementia, no dementia, and very mild dementia.
Performance validation outcomes.
| Classes | Accuracy | Precision | Recall | F1-score | Specificity | IoU |
|---|---|---|---|---|---|---|
| Mild dementia | 98.5 | 98.9 | 97.8 | 98.35 | 99.8 | 98.4 |
| Moderate dementia | 99.3 | 97.9 | 98.3 | 98.1 | 98.8 | 98.6 |
| Non-demented | 99.4 | 98.6 | 98.6 | 98.6 | 98.7 | 99.3 |
| Very mild dementia | 99.8 | 99.1 | 99.1 | 99.1 | 99.3 | 99.6 |
Abbreviation: IoU, intersection over union.
The performance evaluation conducted on the OASIS dataset offers vital insights into the model’s capacity to appropriately categorize people at different phases of dementia development. The model’s ability to accurately differentiate between various dementia subtypes and identify people with cognitive impairment from those without is confirmed by the high level of accuracy achieved on this dataset.
The model’s capacity to precisely categorize patients with mild, moderate, and very mild dementia underscores its sensitivity to modest variations in cognitive performance and the extent of the illness. The model successfully detects structural and functional problems related to various phases of AD development by accurately collecting subtle patterns and characteristics from MRI scans.
Furthermore, it is essential to include non-demented persons as a distinct category in the classification task to evaluate the model’s specificity and capacity to distinguish between healthy individuals and those with cognitive impairment. The model’s ability to accurately categorize non-demented people highlights its dependability in differentiating AD-related pathology from normal brain morphology.
The comparative analysis outcomes are revealed in Table 2. Similarly, Figure 3 highlights the significance of the proposed model. The SVM model demonstrated an impressive average accuracy of 99.3% in multi-class classification, demonstrating the strength and effectiveness of the suggested model for detecting AD. In order to confirm its effectiveness, the authors conducted a comparison of the model with various backbones, such as EfficientNet B7, MobileNet V3, and models proposed by Nancy Noella and Priyadarshini (2023), An et al. (2020), Choi and Lee (2020), and Lella et al. (2021). This evaluation was performed using the OASIS dataset, which includes classes representing different levels of dementia severity: mild dementia, moderate dementia, non-demented, and very mild dementia.
Comparative analysis outcomes.
| Model | Accuracy | Precision | Recall | F1-score | Specificity | IoU |
|---|---|---|---|---|---|---|
| Proposed model | 99.3 | 98.6 | 98.5 | 98.5 | 99.2 | 99.0 |
| Nancy Noella and Priyadarshini (2023) | 96.8 | 97.5 | 97.9 | 97.7 | 97.8 | 96.4 |
| An et al. (2020) | 98.4 | 98.1 | 98.3 | 98.2 | 98.2 | 98.7 |
| Choi and Lee (2020) | 97.8 | 96.8 | 97.6 | 97.2 | 96.4 | 95.9 |
| Lella et al. (2021) | 98.6 | 97.5 | 96.8 | 97.1 | 94.8 | 96.7 |
| EfficientNet B7 | 97.6 | 96.7 | 97.8 | 97.2 | 96.5 | 94.8 |
| MobileNet V3 | 98.1 | 97.8 | 98.6 | 98.2 | 97.5 | 93.8 |
Abbreviation: IoU, intersection over union.
The presented model outperformed the other models in all classes, reaching a much greater accuracy. Our model properly categorizes patients with mild, moderate, and very mild dementia, and non-demented status. This demonstrates its sensitivity to even slight changes in cognitive function and disease severity, surpassing the performance of existing models in this regard.
The comparison also includes EfficientNet B7 and MobileNet V3, which are renowned for their effectiveness and high performance in activities related to picture categorization. Although these models have shown encouraging outcomes in other fields, our EL method, which incorporates feature extraction using YOLOv7 and EfficientNet B3, has proven to be more efficient in detecting AD, especially in discriminating between different stages of dementia.
In addition, while comparing the proposed model to the models suggested by Nancy Noella and Priyadarshini (2023), An et al. (2020), Choi and Lee (2020), and Lella et al. (2021), it was consistently found that our model performed better in terms of classification accuracy on the OASIS dataset. The results indicate that the combination of EL, feature extraction from YOLOv7, and EfficientNet B3 outperforms previous approaches presented in the literature for AD detection.
Overall, the performance validation conducted on the OASIS dataset highlights the efficacy of our proposed model in accurately classifying AD across various stages of dementia severity. The exceptional precision attained in comparison to other models underscores the potential of our method as a helpful diagnostic instrument for precisely categorizing people according to their cognitive state and illness advancement. Nevertheless, further verification using more extensive and more varied datasets, clinical evaluation, and longitudinal investigations would be imperative to substantiate the model’s clinical efficacy and applicability in real-world scenarios.
Table 3 presents the computational costs of each AD identification model. The suggested model demonstrated exceptional multi-class classification performance on the OASIS dataset, with notable model parameters of 27 millions and floating point operations (FLOPs) of 35 giga. In order to conduct a thorough comparison, we assessed our model against other advanced frameworks, such as EfficientNet B7, MobileNet V3, and models presented by Nancy Noella and Priyadarshini (2023), An et al. (2020), Choi and Lee (2020), and Lella et al. (2021).
Computational complexities.
| Model | Parameters (in millions) | FLOPs (in giga) | Learning rate | Testing time (seconds) |
|---|---|---|---|---|
| Proposed model | 27 | 35 | 0.0001 | 87.5 |
| Nancy Noella and Priyadarshini (2023) | 52 | 67 | 0.005 | 103.2 |
| An et al. (2020) | 48 | 67 | 0.004 | 118.8 |
| Choi and Lee (2020) | 56 | 75 | 0.004 | 98.7 |
| Lella et al. (2021) | 47 | 65 | 0.0001 | 125.23 |
| EfficientNet B7 | 38 | 57 | 0.008 | 175.1 |
| MobileNet V3 | 45 | 62 | 0.0001 | 109.8 |
Abbreviation: FLOPs, Floating point operations.
EfficientNet B7 and MobileNet V3, well-known for their high efficiency and efficacy in picture classification tasks, were used as robust benchmarks in our comparison. Although they demonstrated impressive architectural abilities, the suggested model surpassed them by a wide margin in terms of classification accuracy on the OASIS dataset. These findings indicate that the proposed EL method, which combines feature extraction from YOLOv7 and EfficientNet B3, outperforms EfficientNet B7 and MobileNet V3 in detecting AD. Specifically, our methodology excels in accurately discriminating between various stages of dementia.
In addition, while comparing our model to the models presented by Nancy Noella and Priyadarshini (2023), An et al. (2020), Choi and Lee (2020), and Lella et al. (2021), it was consistently found that our model achieved greater classification accuracy for all classes. This highlights the effectiveness of our methodology in detecting AD compared to the approaches presented in previous literature.
The exceptional success of our model may be credited to several elements, such as the harmonious combination of YOLOv7 and EfficientNet B3 for extracting features, the use of EL framework, and the careful tuning of model parameters. Our technique exhibited higher performance in AD classification tasks, showcasing its efficiency and efficacy while having fewer parameters and lower computing complexity than previous models.
The model’s performance should be validated on diverse populations to ensure its effectiveness across different demographic groups. Variations in demographics, genetics, and environmental factors can impact disease manifestation and progression. Demonstrating the model’s performance in real-world clinical settings, including its impact on patient outcomes and clinical decision-making, is essential for assessing its practical utility and value. YOLOv7 and EfficientNet B3 can be computationally intensive, requiring significant resources for training and inference. Scaling the models to handle large datasets efficiently can be challenging.
Training SVM can be a relatively time-consuming process, especially with large datasets and high-dimensional feature spaces. The computational complexity of SVM training scales quadratically with the number of training samples, making it less suitable for large-scale datasets. When deploying the ensemble model in real-world clinical settings, it’ is essential to consider the computational resources available and optimize the model architecture accordingly for efficient inference on target hardware platforms.
Ensuring that the proposed ensemble architecture can scale effectively to handle increased data volumes and maintain performance is essential for its practical applicability in clinical settings. Integrating AI-driven diagnostic tools into existing clinical workflows presents logistical and regulatory challenges. Ensuring seamless integration and compliance with healthcare standards and regulations is essential for widespread adoption.
The proposed diagnostic instrument for AD detection using MRI data offers promising opportunities to enhance clinical practice, facilitate early intervention strategies, and improve patient care and outcomes. By complementing existing clinical practices with objective and reliable measures of disease progression, the instrument empowers healthcare professionals to deliver personalized, timely, and effective care to individuals affected by AD. Furthermore, its potential to accelerate research efforts and reduce the overall burden of the disease underscores its significance in advancing the field of AD diagnosis and management.
Exploring advanced DL architectures such as three-dimensional (3D) CNN and graph convolutional networks designed for processing 3D volumetric data can more effectively capture spatial relationships within the brain. Leveraging self-supervised learning techniques for pre-training DL models on unlabeled MRI data can improve generalization and feature representation, especially in scenarios with limited labeled data. Including molecular biomarkers such as CSF levels of amyloid-beta and tau proteins can improve the accuracy of AD detection. Molecular biomarkers offer insights into the underlying pathological processes of AD and can serve as valuable indicators for early diagnosis and disease monitoring. Integrating genetic markers associated with AD risk, such as Apolipoprotein E genotype, into the model can enhance risk prediction and stratification, enabling personalized interventions and treatments.
CONCLUSIONS
The EL-based AD detection technique using YOLOv7, EfficientNet B3, CatBoost, XGBoost, and SVM models shows promise for improving diagnostic accuracy and resilience. Using cutting-edge DL methods like YOLOv7 and EfficientNet B3, researchers can extract features from MRI images with great precision. This allows us to capture spatial and semantic information, resulting in a more complete picture of AD pathology. By including CatBoost and XGBoost as base learners and SVM as meta-learners, the model integrated different predictions and improved its capacity to differentiate between AD and healthy brain images. This ensemble approach enhanced both the accuracy of classification and the ability of the model to generalize and remain resilient, which are essential characteristics in real-world applications of AD detection. EL enables us to use the distinct capabilities of each model component, leading to a synergistic approach that surpasses the performance of individual classifiers. Moreover, the exceptional precision attained on the Kaggle repository dataset highlights the efficacy of our approach in detecting AD. In order to assess the scalability, generalizability, and clinical application of the suggested EL technique, further research is necessary in the future. Long-term investigations and verification on varied datasets are essential to evaluate the model’s effectiveness on various demographics and imaging systems. Furthermore, it is necessary to make significant efforts to improve the comprehensibility and clarity of the model’s predictions in order to facilitate its acceptance and use in clinical settings. In summary, the framework for detecting AD using EL has great potential for improving early diagnostic and therapeutic methods.