Discriminative Sparse Features for Alzheimer's Disease Diagnosis Using Multimodal Image Data

Area of Science:

• Medical Image Analysis
• Machine Learning
• Neuroscience

Background:

• High-dimensional medical imaging datasets pose challenges for feature extraction due to the curse of dimensionality.
• Limited sample sizes in medical datasets exacerbate the difficulty of extracting discriminative features for classification tasks.

Purpose of the Study:

• To develop an effective method for feature extraction and classification in high-dimensional medical imaging data.
• To address the challenge of limited samples by utilizing sparse representations for robust feature encoding.
• To enable multimodal image classification by combining specialized classifiers for PET and MRI data.

Main Methods:

• Employed sparse representations for data analysis and feature extraction.
• Developed a novel method to combine Support Vector Classifiers (SVC) by leveraging distance to hyperplane for modality selection.
• Utilized multimodal data (PET and MRI) for classification of Alzheimer's Disease (AD), Mild Cognitive Impairment (MCI), and Control (CN) subjects.

Main Results:

• Achieved classification accuracies of up to 92% for CN/AD and 84% for CN/MCI.
• Demonstrated the effectiveness of the proposed method through cross-validation experiments on a dataset from the Alzheimer's Disease Neuroimaging Initiative.
• Showcased differential discriminative power between imaging modalities for different disease stages.

Conclusions:

• Sparse representations are crucial for encoding and transferring salient image features efficiently.
• The proposed method outperforms traditional projection techniques like Principal Component Analysis (PCA) in feature extraction for medical images.
• The approach provides insights into disease progression, highlighting functional changes in AD and structural changes in MCI.