Voxel-Wise Feature Selection Method for CNN Binary Classification of Neuroimaging Data.

Area of Science:

• Neuroimaging
• Machine Learning
• Artificial Intelligence

Background:

• Deep learning (DL) models require effective feature selection (FS) for brain imaging data classification.
• Traditional FS methods may not optimally leverage the spatial information inherent in neuroimaging datasets.

Purpose of the Study:

• To introduce and evaluate a novel voxel-wise group analysis technique, t-masking, as a data-driven FS strategy for DL in brain imaging.
• To assess the impact of t-masking on the classification performance of a convolutional neural network (CNN) for Alzheimer's disease detection.

Main Methods:

• Implemented t-masking, a FS technique based on voxel-wise two-sample t-tests, integrated within a CNN learning procedure.
• Utilized a structural magnetic resonance imaging dataset of 180 subjects for binary classification of very-mild Alzheimer's disease versus normal controls.
• Designed six experimental configurations to analyze the impact of t-masking and compared its performance against other FS-based machine learning (ML) models.

Main Results:

• The t-masking approach demonstrated an approximate 6% enhancement in classification performance.
• The performance improvement achieved with t-masking was superior to that of comparable ML models employing different FS strategies.
• Evaluated the influence of t-masking on various selection rates, providing insights for future research.

Conclusions:

• Voxel-wise group analysis via t-masking offers a significant improvement for DL-based brain imaging classification tasks.
• The t-masking method shows high generalizability across different DL architectures, neuroimaging modalities, and various brain pathologies.
• This data-driven FS strategy enhances diagnostic accuracy and offers a valuable tool for neurodegenerative disease research.