EARLY ALZHEIMER’S DISEASE DETECTION USING A HYBRID DEEP LEARNING APPROACH ON MRI IMAGING

ABSTRACT

This study presents a deep learning-based method for early detection of Alzheimer’s Disease (AD) using structural MRI data. A hybrid architecture, combining 3D Convolutional Neural Networks (CNNs), temporal attention, and Recurrent Neural Networks (RNNs), was trained and evaluated on the OASIS-2 dataset. This dataset contains longitudinal T1-weighted MRI scans labeled with Clinical Dementia Rating (CDR) scores. The model demonstrated strong performance, particularly in detecting early-stage AD, achieving an F1-score of 0.92 and an overall accuracy of 86%. Attention mechanisms contributed to the interpretability of results, highlighting key brain areas involved in disease progression. These findings suggest that integrating spatial and temporal modeling can significantly improve diagnostic accuracy and provide a scalable approach for real-world screening scenarios.

АННОТАЦИЯ

В данной работе предложен метод ранней диагностики болезни Альцгеймера (БА), основанный на использовании структурных МРТ-данных и гибридной модели глубокого обучения. Архитектура модели включает в себя сверточные нейронные сети (3D CNN), механизм временного внимания и рекуррентные нейронные сети (RNN). Обучение и оценка производились на наборе данных OASIS-2, содержащем МРТ-сканы головного мозга, размеченные по шкале клинической деменции (CDR). Модель показала высокую точность в выявлении ранней стадии БА (F1-мера — 0.92) и общую точность классификации 86%. Механизм внимания позволил визуализировать наиболее значимые области мозга, участвующие в развитии заболевания. Полученные результаты демонстрируют потенциал использования комбинированных пространственно-временных моделей для повышения точности диагностики и практического применения в клинической среде.

Keywords: Alzheimer’s Disease, early diagnosis, structural MRI, deep learning, temporal attention, CNN, RNN, medical imaging.

Ключевые слова: Болезнь Альцгеймера, ранняя диагностика, структурная МРТ, глубокое обучение, временное внимание, CNN, RNN, медицинская визуализация.

Introduction

Alzheimer’s Disease is one of the most common forms of dementia, affecting millions of people worldwide. It leads to progressive decline in memory, reasoning, and behavioral abilities, eventually resulting in the loss of independent function. According to recent projections, the number of individuals living with AD is expected to exceed 100 million globally by 2050, placing a significant burden on healthcare systems and families [1], [7].

Although AD is irreversible, early detection can greatly improve treatment outcomes. It allows patients to access medications and interventions that may slow the progression of symptoms. Moreover, early diagnosis gives families time to plan and make important lifestyle and care-related decisions. However, traditional diagnostic methods, such as cognitive tests and clinical observations, often fail to detect AD in its earliest stages.

To address this limitation, researchers have turned to neuroimaging techniques, especially magnetic resonance imaging (MRI), which enables visualization of the brain’s structural changes. Studies have shown that brain regions such as the hippocampus shrink noticeably in the early stages of AD [2], [4]. With the help of machine learning (ML) and deep learning (DL) models, it is now possible to detect these subtle alterations automatically, offering a promising approach for early diagnosis.

Deep learning architectures like convolutional neural networks (CNNs) have demonstrated strong performance in analyzing medical images. Furthermore, when combined with attention mechanisms and recurrent layers, such models can not only classify, but also highlight areas of diagnostic relevance [6], [15]. This increases both interpretability and clinical trust.

In this paper, we use a hybrid model that includes CNNs, attention mechanisms, and RNNs. These components help the model focus on relevant time sequences and spatial brain regions, improving its ability to classify patients accurately. Alzheimer’s Disease is a progressive condition that mainly affects memory and cognitive abilities. As the population ages, the number of AD cases is expected to increase dramatically in the coming decades [7]. Detecting the disease early can allow for better planning, early treatment, and improved life quality for patients and their families [1]. However, diagnosing AD at an early stage is still difficult due to overlapping symptoms with other conditions.

Recent advances in brain imaging and artificial intelligence have helped create new tools for early diagnosis. Among these, MRI is a safe and non-invasive way to observe structural changes in the brain [2], [6]. When MRI data is combined with deep learning models, especially those using CNNs, it becomes possible to detect subtle changes that may signal early AD.

In this paper, we use a hybrid model that includes CNNs, attention mechanisms, and RNNs. These components help the model focus on relevant time sequences and spatial brain regions, improving its ability to classify patients accurately.

Materials and methods

Model architecture

Figure 1 shows the hybrid deep learning model used in this work. The model starts with a 3D CNN that extracts spatial patterns from MRI volumes. Then, a temporal attention layer weighs the significance of different slices. The output is passed to a recurrent neural network (RNN), which captures time-based relationships across the scan sequence. Finally, the classification layer predicts the patient’s cognitive stage.

Figure 1. Hybrid Model Architecture

Figure 2 demonstrates the preparation of the MRI dataset. Each subject’s scan is preprocessed by removing non-brain tissue, followed by spatial normalization and motion correction. The 3D structural images are classified according to CDR scores, enabling the early detection of Alzheimer's disease.

Figure 2. MRI Dataset Processing Pipeline

Dataset

The OASIS-2 dataset contains longitudinal MRI scans of elderly subjects, captured over multiple sessions. Each subject is labeled using Clinical Dementia Rating (CDR) scores. Subjects fall into four groups: Healthy (CDR=0), MCI-1 (CDR=0.5), MCI-2 (CDR=1), and Early AD (CDR≥2) [2]. All scans go through skull stripping, spatial alignment, and intensity correction. This dataset is widely used in research because of its quality and open access [10].

Evaluation And Metrics

To evaluate the model, we used standard classification metrics: precision, recall, accuracy, and F1-score. A 5-fold cross-validation method was applied to check the stability of the results. Data augmentation techniques like affine transformation and temporal windowing were used to improve performance.

The classification report is shown in Table 1. The model performed best in identifying Early AD cases (F1-score: 0.92), while classification of MCI stages was more challenging due to feature overlap and fewer training examples.

Table 1.

Classification Report for AD Staging

Results And Discussions

The high performance of the model in detecting Early AD suggests that combining spatial and temporal information is beneficial [6], [14]. Attention maps showed strong activation in areas such as the hippocampus and default mode network, which are commonly associated with AD progression [4], [5].

Compared to traditional models like ResNet and Naive Bayes, the hybrid approach used here provides better balance between accuracy and interpretability [3], [12]. Similar strategies using attention-based architectures have shown promise in previous studies [6], [15].

Conclusion

This study demonstrates that deep learning models using structural MRI data can effectively detect early signs of Alzheimer’s Disease. By combining CNNs, attention mechanisms, and RNNs, the proposed hybrid model is able to capture both spatial and temporal features from the brain, which results in significantly improved classification performance.

The achieved accuracy of 86% and high F1-score for early AD detection highlight the robustness and reliability of the model. This level of performance makes the approach suitable for use in real-world clinical scenarios, where timely and accurate diagnosis is critical. Moreover, the model’s interpretability through attention mapping adds an additional layer of trust and transparency for medical practitioners.

This work also reinforces the growing consensus that automated systems based on deep learning can play an essential role in the early detection of neurodegenerative diseases. It opens up opportunities to integrate AI into routine screening programs, particularly in regions with limited access to expert neurologists or radiologists.

However, there are still challenges to address. For instance, variations in imaging protocols, demographic differences, and data imbalance can affect generalization. Future research should aim to validate the model across diverse populations and imaging conditions. Additionally, combining structural MRI with other modalities such as PET, EEG, or fluid biomarkers may further enhance diagnostic accuracy and provide a more holistic view of the disease progression [8], [14].

Overall, this study takes a promising step toward more accessible, efficient, and explainable diagnostic tools for Alzheimer’s Disease that could transform the way early detection is approached in both clinical and research settings.

References:

1. C. R. Jack et al., “Revised criteria for diagnosis and staging of Alzheimer’s disease: Alzheimer’s Association Workgroup,” Alzheimer’s & Dementia, vol. 20, no. 8, pp. 5143–5169, 2024.
2. M. C. Carrillo et al., “The role of the Alzheimer’s Association in the genesis of Alzheimer’s Disease Neuroimaging Initiative,” Alzheimer’s and Dementia, 2024.
3. A. Chandra and S. Roy, “On the Detection of Alzheimer’s Disease using Naïve Bayes Classifier,” in Proc. ICMOCE, 2023.
4. Q. Zhang et al., “Brain Connectome Imaging Markers Research of Glucose Metabolism in the Early Stage of Alzheimer’s Disease,” in Proc. IEEE EMBS, 2023.
5. U. M. R. Munipalli and V. Annepu, “Prediction and Classification of Alzheimer’s Disease Using RESNET-152V2,” in Proc. ISED, 2024.
6. Y. Wei et al., “A deep spatio-temporal attention model of dynamic functional network connectivity shows sensitivity to Alzheimer’s in asymptomatic individuals,” in Proc. IEEE EMBC, 2024.
7. M. Pedersen et al., “Multilayer network switching rate predicts brain performance,” PNAS, vol. 115, no. 52, pp. 13376–13381, 2018.
8. R. Shi et al., “An unsupervised region of interest extraction model for tau PET images and its application in the diagnosis of Alzheimer’s disease,” in Proc. IEEE EMBC, 2022.
9. M. Dixit et al., ”Comparative Analysis of Deep Learning and Machine Learning Models for Alzheimer’s and Parkinson’s Disease Classification from Medical Images,” in Proc. Int. Conf. on Electrical, Electronics and Computing Technologies (ICEECT), 2024.
10.  A. D. Arya et al., “A systematic review on machine learning and deep learning techniques in the effective diagnosis of Alzheimer’s disease,” Brain Informatics, vol. 10, 2023.
11.  A. Khan and S. Zubair, “A Machine Learning-based robust approach to identify Dementia progression employing Dimensionality Reduction in Cross-Sectional MRI data,” in Proc. SMART-TECH, pp. 237–242, 2020.
12. D. R. Khadatkar and J. P. Patra, “Comparative Analysis of Different Machine Learning Algorithms for Detection of Alzheimer Disease from Medical Images,” in Proc. ICAIIHI, 2023.
13. Y. Subbarayudu and A. Sureshbabu, “A Distributed Densely Connected Convolutional Network Approach for Enhanced Recognition of Health-Related Topics: A Societal Analysis Case Study,” Ingenierie des Systemes d’Information, vol. 28, no. 3, pp. 677–684, 2023.
14. N. Sharma, M. H. Kolekar, and K. Jha, “Iterative Filtering Decomposition Based Early Dementia Diagnosis Using EEG with Cognitive Tests,” IEEE Trans. Neural Syst. Rehabil. Eng., vol. 28, no. 9, pp. 1890–1898, 2020.
15. S. Luz, F. Haider, and P. De Sousa, “Machine Learning models for detection and assessment of progression in Alzheimer’s disease based on blood and cerebrospinal fluid biomarkers,” in Proc. IEEE EMBC, 2023.