SF2Former: Amyotrophic lateral sclerosis identification from multi-center MRI data using spatial and frequency fusion...

Area of Science:

• Neuroimaging
• Artificial Intelligence
• Biomedical Engineering

Background:

• Amyotrophic Lateral Sclerosis (ALS) is a neurodegenerative disease impacting motor neurons.
• Brain Magnetic Resonance Imaging (MRI) shows potential as a biomarker for ALS diagnosis and monitoring.
• Current deep learning methods struggle to accurately classify ALS patients from controls due to subtle neuroimaging changes.

Purpose of the Study:

• To develop an advanced deep learning framework, SF²Former, for improved classification of ALS patients.
• To leverage the vision transformer architecture for identifying long-range relationships in neuroimaging data.
• To enhance classification performance by integrating spatial and frequency domain information from MRI scans.

Main Methods:

• The SF²Former framework utilizes a vision transformer architecture to analyze MRI data.
• It combines spatial and frequency domain information for enhanced feature representation.
• The model is trained on consecutive coronal MRI slices using transfer learning from ImageNet and employs majority voting for final classification.

Main Results:

• The proposed SF²Former framework demonstrated superior classification accuracy in distinguishing ALS patients from healthy controls.
• Multi-modal neuroimaging data (T1-weighted, R2*, FLAIR) from the CALSNIC datasets were used for evaluation.
• The framework outperformed several popular deep learning-based techniques in experimental results.

Conclusions:

• SF²Former offers a promising approach for accurate ALS diagnosis using neuroimaging data.
• The integration of spatial and frequency domain analysis significantly improves deep learning model performance.
• This framework has the potential to advance the clinical application of MRI in ALS management.