Scientific research project

Abstract

Alzheimer's disease stands as the most prevalent form of dementia, recognized as one of the foremost causes of mortality globally. Timely diagnosis opens avenues to accessible treatment modalities, substantially ameliorating cognition and elevating life quality. Presently, MRI brain imaging emerges as a pivotal tool for early Alzheimer's detection and diagnosis. This study introduces a novel classification methodology leveraging the architectures of InceptionV3, Xception, and Vision Transformer networks. Additionally, it proposes three innovative network models derived from the Vision Transformer model ResNet50, EfficientNetV2, and VGG16 to categorize Alzheimer's disease images into normal, mild, very mild, or moderate states. Moreover, the research incorporates the YOLOV8 network for anomaly region identification. Experimental outcomes underscore the efficacy of the proposed models, achieving an impressive 99.45% accuracy on the OASIS MRI dataset and 96.6% on the Alzheimer's MRI dataset.

Introduction

Alzheimer's disease (AD) is the leading cause of dementia, affecting over 55 million people globally as of 2019, with predictions reaching 78 million by 2030. AD accounts for 60–70% of dementia cases and is a major cause of death, particularly among women. Despite its prevalence, over 50% of dementia patients remain undiagnosed, although most express a desire for early diagnosis. AD is characterized by progressive neurodegeneration, marked by beta-amyloid plaques, neurofibrillary tangles, tau proteins, and reactive gliosis. Early symptoms are often subtle, complicating timely diagnosis and treatment. Early detection, before clinical symptoms appear, is crucial for effective intervention. While imaging techniques like MRI, PET, and CT can detect early brain changes in AD, MRI is the preferred method due to its effectiveness and safety, avoiding the high costs and harmful radiation associated with PET and CT. Given the importance of early diagnosis, I have chosen to research "Detection and diagnosis of Alzheimer's disease using the Vision Transformer technique" to aid in early detection and improve patient outcomes.

The objectives of the project include gaining knowledge about embedded programming, sensors, functional modules, and wireless communication methods. The model will be designed to have good expandability, with standardized, easily classified, and user-friendly connectors. The project also focuses on studying popular microcontrollers such as ESP32, ESP8266, Arduino, and Raspberry Pi, along with the application of the C++ programming language. The practical significance of the project not only helps students grasp IoT knowledge but also lays the foundation for further research and IoT applications.

Figure 1. Brain MRI images of the stages of Alzheimer's disease

Proposed Method

This study uses a transfer learning method based on deep learning techniques and the Vision Transformer network to detect and classify AD diseases. Details of the stages are shown in Figure 2. The illustration model proposes the detection and classification of AD disease, after having MRI image data, the data set is processed and divided into 2 train and test data sets, each data set will have 4 layers: Mild Demented, Moderate Demented, Non Demented, Very Mild Demented. After being divided and processed, the data will be fed into the model to extract characteristics and train to produce the model and suitable parameters for AD detection and classification

Figure 2. Proposed model for the detection and classification of Alzheimer's disease

To conduct an experiment for the proposed model, this study implemented 6 scenarios on two proposed datasets. Specific parameters in Table 1

Table 1. Recommended Scenarios and Training Parameters

Results

Figure 3 presents some taxonomic images of the proposed scenarios on the Alzheimer's MRI dataset. Scenario 1 correctly predicts 3 actual labels: Non Demented, Very Mild Demented, Mild Demented, scenario 2 correctly predicts 2 labels Non Demented, Very Mild Demented. Scenario 6 also only correctly predicted 2 labels: Non Demented, Very Mild Demented. Scenarios 3, 4, and 5 correctly predict all labels. However, scenario 3 has the highest accuracy of all scenarios. The results show that scenario 3 works much better than the other 5 scenarios.

Figure 3. Some classification results on the Alzheimer's MRI dataset

Figure 4 presents some taxonomic images of the scenarios that have been proposed on the OASIS MRI dataset. The results showed that all 6 scenarios correctly predicted all 4 labels that were true to the reality of Non Demented, Very Mild Demented, Mild Demented and Moderate Demented. In which, scenarios 3, 4, and 5 correctly predict all labels with 100% accuracy. This showed that the predictive models were better on the OASIS MRI dataset than on the Alzheimer's MRI dataset

Figure 4. Some images of the classification results on the test dataset OASIS MRI

Conclusions

Detecting and accurately diagnosing Alzheimer's disease is a significant challenge due to its high fatality rate and progressive nature. Early detection is crucial for timely treatment. This study proposes using deep learning models, such as InceptionV3, Xception, Vision Transformer, VGG16, EfficientNetV2S, and ResNet50, for Alzheimer's disease classification to evaluate model performance and enhance prediction accuracy. Additionally, the YoloV8 model is employed for localizing abnormalities in the brains of Alzheimer's patients. Experimental results demonstrate that the Vision Transformer network achieves superior accuracy, with up to 96.6% on the Alzheimer's MRI dataset and 99.45% on the OASIS MRI dataset. These findings contribute to developing a system for detecting and evaluating Alzheimer's disease, aiding doctors in making accurate diagnoses and providing timely treatment.

References