Alzheimer's and Parkinson's Detection with Video-Based Hybrid Deep Learning from Brain MRI.
This study applies CNN + LSTM/GRU models to MRI-derived short videos to classify Alzheimer's, Parkinson's, and controls, reporting very high accuracy (~99.7%) and providing a GUI for deployment.
What the AI sees
This study applies CNN + LSTM/GRU models to MRI-derived short videos to classify Alzheimer's, Parkinson's, and controls, reporting very high accuracy (~99.7%) and providing a GUI for deployment.
Research significance
Has limited direct therapeutic discovery value—it's primarily a diagnostic/classification tool that could aid patient stratification or trial recruitment if rigorously validated, but it offers no mechanistic insights or intervention targets for Parkinson's disease.
Source abstract
Dementia, which refers to disorders related to human memory, significantly affects the human brain, and a person with it can experience certain difficulties in physical and mental activities. The most common and fatal types of dementia are Alzheimer's Disease (AD) and Parkinson's Disease (PD). Therefore, AD, PD, and control labels are detected in this study from videos created using brain Magnetic Resonance Imaging (MRI). A public dataset including these three classes is used in this study. After preprocessing this dataset, a video is created for each class. Afterward, short video clips are randomly obtained from these videos. These video clips are randomly split into 50% training and 50% validation sets. Herein, features are extracted for training and validation using a Convolutional Neural Network (CNN)-based architecture. Long Short-Term Memory (LSTM), LSTM + Gated Recurrent Units (GRU), and Deeper LSTM architectures are trained using these extracted features. In addition, a user-friendly Graphical User Interface (GUI) application including all three models developed in the study is designed for AD, PD, and control detection. It is noted that these video-based architectures can achieve high performance with fewer short video clips, although 50% of the training data is used in the study. The maximum accuracy and specificity values achievable by the models developed in the study are 99.67% and 99.83%, respectively.