Efficient deep learning models for early Alzheimer’s disease diagnosis
Abstract
Alzheimer’s Disease (AD) is a progressive neurological disorder that impairs cognitive functions, severely affecting patients’ quality of life. Thus, early and accurate diagnosis is essential to provide a chance for treatment. This paper proposes two deep learning (DL) models to enable automated diagnosis of medical images. The first model is a hybrid of a convolutional neural network and bidirectional long short-term memory (CNN-BiLSTM) for classification of AD, and the second model is a U-Net for a segmentation task. The CNN-BiLSTM is applied to an overall dataset that consists of 6400 magnetic resonance imaging (MRI) images, while the U-Net is evaluated on a dataset that contains hippocampus regions that are isolated from the MRI images with manually labelled ground truth masks. The CNN-BiLSTM is implemented from scratch to detect four classes of AD: Non-Demented (ND), Mild Demented (MD), Moderate Demented (MOD), and Very Mild Demented (VMD). To tackle the issue of imbalanced classes in the dataset, data augmentation methods are employed to balance the dataset. The proposed CNN-BiLSTM model incorporates a Gradient-weighted Class Activation Mapping (Grad-CAM) processing technique to offer visual interpretations for its predictions. The Grad-CAM highlights the most relevant regions in MRI images that influence the model’s decisions and offers interpretability. Also, three explainable AI (XAI) techniques are applied to interpret the decision of the CNN-BiLSTM in the classification: Grad-CAM++, integrated gradient, and saliency map. The efficiency of the U-Net is assessed using two measurements: Dice Score and Intersection over Union (IoU). The proposed U-Net model achieved a Dice score of 92% and a mean IoU of 86%, which indicates high segmentation accuracy, while the CNN-BiLSTM achieves superior efficiency, with a testing accuracy of 99% and a testing loss rate of 0.0408. Also, the CNN-BiLSTM is measured by Precision, Recall, and F1-score metrics, all reaching 99%. The CNN-BiLSTM attains areas under the receiver operating characteristic (ROC) curves of 100%. These results underscore the potential of the proposed CNN-BiLSTM and U-Net integration to enhance the accuracy and reliability for diagnosing AD.
Related
- A general practice-centred model for early diagnosis and management of Alzheimer's disease in Denmark: rationale, evidence and a pragmatic diagnostic approach.
- Multimodal patch-based deep learning for early Alzheimer’s disease classification using MRI and PET
- An effective and deep learning model to predict Alzheimer’s disease at an early stage from audio data
- Efficient Alzheimer’s disease classification with robust rank pooling
- Amyloid PET Quantitation and Centiloid Thresholds in the Diagnosis of Alzheimer Disease
BibTeX
@article{oraby2026efficient,
title = {{Efficient deep learning models for early Alzheimer’s disease diagnosis}},
author = {Sarah Oraby and Ahmed Emran and Basel El-Saghir and Saeed Mohsen},
journal = {Scientific Reports},
year = {2026},
month = oct,
doi = {10.1038/s41598-026-71591-2},
url = {https://www.nature.com/articles/s41598-026-71591-2},
note = {CC BY 4.0},
}
RIS
TY - JOUR
TI - Efficient deep learning models for early Alzheimer’s disease diagnosis
AU - Sarah Oraby
AU - Ahmed Emran
AU - Basel El-Saghir
AU - Saeed Mohsen
PY - 2026
DA - 2026/10/06
JO - Scientific Reports
DO - 10.1038/s41598-026-71591-2
UR - https://www.nature.com/articles/s41598-026-71591-2
AB - Alzheimer’s Disease (AD) is a progressive neurological disorder that impairs cognitive functions, severely affecting patients’ quality of life. Thus, early and accurate diagnosis is essential to provide a chance for treatment. This paper proposes two deep learning (DL) models to enable automated diagnosis of medical images. The first model is a hybrid of a convolutional neural network and bidirectional long short-term memory (CNN-BiLSTM) for classification of AD, and the second model is a U-Net for a segmentation task. The CNN-BiLSTM is applied to an overall dataset that consists of 6400 magnetic resonance imaging (MRI) images, while the U-Net is evaluated on a dataset that contains hippocampus regions that are isolated from the MRI images with manually labelled ground truth masks. The CNN-BiLSTM is implemented from scratch to detect four classes of AD: Non-Demented (ND), Mild Demented (MD), Moderate Demented (MOD), and Very Mild Demented (VMD). To tackle the issue of imbalanced classes in the dataset, data augmentation methods are employed to balance the dataset. The proposed CNN-BiLSTM model incorporates a Gradient-weighted Class Activation Mapping (Grad-CAM) processing technique to offer visual interpretations for its predictions. The Grad-CAM highlights the most relevant regions in MRI images that influence the model’s decisions and offers interpretability. Also, three explainable AI (XAI) techniques are applied to interpret the decision of the CNN-BiLSTM in the classification: Grad-CAM++, integrated gradient, and saliency map. The efficiency of the U-Net is assessed using two measurements: Dice Score and Intersection over Union (IoU). The proposed U-Net model achieved a Dice score of 92% and a mean IoU of 86%, which indicates high segmentation accuracy, while the CNN-BiLSTM achieves superior efficiency, with a testing accuracy of 99% and a testing loss rate of 0.0408. Also, the CNN-BiLSTM is measured by Precision, Recall, and F1-score metrics, all reaching 99%. The CNN-BiLSTM attains areas under the receiver operating characteristic (ROC) curves of 100%. These results underscore the potential of the proposed CNN-BiLSTM and U-Net integration to enhance the accuracy and reliability for diagnosing AD.
SN - 2045-2322
ER -