A robust class decomposition-based approach for detecting Alzheimer's progression.

Area of Science:

• Neuroimaging
• Artificial Intelligence
• Medical Diagnostics

Background:

• Computer-aided diagnosis of Alzheimer's disease (AD) using structural magnetic resonance imaging (sMRI) is advancing.
• Training deep learning models like convolutional neural networks from scratch is hindered by extensive data and computational requirements.
• Class imbalance in datasets can lead to poor predictive performance in machine learning models.

Purpose of the Study:

• To propose and evaluate a class decomposition transfer learning (CDTL) approach for detecting AD from sMRI.
• To assess the robustness of the CDTL approach across various Alzheimer's Disease Neuroimaging Initiative (ADNI) cohorts.
• To improve the prediction of mild cognitive impairment (MCI) to AD conversion.

Main Methods:

• Employed transfer learning by leveraging pre-trained models (VGG19, AlexNet).
• Utilized a class decomposition technique to handle dataset irregularities and class imbalance.
• Integrated an entropy-based method for enhanced classification.

Main Results:

• The CDTL approach demonstrated effectiveness in detecting AD from sMRI data.
• Comparable classification accuracy was observed across different ADNI cohorts, indicating robustness.
• Achieved state-of-the-art performance with 91.45% accuracy in predicting MCI-to-AD conversion.

Conclusions:

• The proposed CDTL method offers a practical and effective solution for AD detection from sMRI.
• Transfer learning and class decomposition are valuable strategies for overcoming challenges in medical image analysis.
• The model shows significant potential for early detection and prediction of AD progression.