Related Experiment Video
Updated: Jun 15, 2025
Author Spotlight: Advancing Alzheimer's Research – Exploring Early Detection and Multi-Omics Approaches
Published on: December 15, 2023
SpectroCVT-Net: A convolutional vision transformer architecture and channel attention for classifying Alzheimer's
Mario Alejandro Bravo-Ortiz1, Ernesto Guevara-Navarro2, Sergio Alejandro Holguín-García1
1Departamento de Electrónica y Automatización, Universidad Autónoma de Manizales, Manizales, Caldas, Colombia; Centro de Bioinformática y Biología Computacional (BIOS), Manizales, Caldas, Colombia.
This study introduces SpectroCVT-Net, a novel AI model for early Alzheimer's disease detection using electroencephalography (EEG) spectrograms. The model achieves high accuracy in distinguishing Alzheimer's, frontotemporal dementia, and healthy controls.
Area of Science:
- Neuroscience
- Artificial Intelligence
- Biomedical Engineering
Background:
- Dementia, including Alzheimer's disease (AD), affects millions globally, necessitating advanced diagnostic tools.
- Current AD diagnosis often relies on specific biomarkers, but early detection remains challenging.
- Electroencephalography (EEG) offers a cost-effective and accessible method for neurological assessment.
Purpose of the Study:
- To develop and evaluate a novel deep learning architecture, SpectroCVT-Net, for classifying dementia subtypes using EEG spectrograms.
- To improve the accuracy and interpretability of AI-driven diagnostic models for neurodegenerative diseases.
- To assess the efficacy of SpectroCVT-Net without relying on transfer learning.
Main Methods:
- EEG signals were processed using Short-Time Fourier Transform (STFT) to create spectrograms.
- A novel Convolutional Vision Transformer (CVT) architecture, SpectroCVT-Net, was designed, integrating convolutional layers with channel attention mechanisms.
- The model was trained and validated on the Brainlat database, including data from Alzheimer's patients, healthy controls, and behavioral variant frontotemporal dementia (bvFTD) patients.
Main Results:
- SpectroCVT-Net achieved a classification accuracy of 92.59 ± 2.3% across Alzheimer's, healthy controls, and bvFTD.
- The model demonstrated superior performance compared to traditional transfer learning methods.
- Grad-CAM analysis provided insights into the network's decision-making process, highlighting critical EEG features.
Conclusions:
- SpectroCVT-Net represents a significant advancement in AI-based EEG analysis for dementia diagnosis.
- The architecture effectively captures both local and global dependencies within EEG spectrograms.
- The study underscores the potential of SpectroCVT-Net to aid in early clinical diagnosis and management of dementia.
