3D scattering transforms for disease classification in neuroimaging.

Area of Science:

• Medical imaging analysis
• Machine learning for healthcare
• Neuroscience research

Background:

• Classifying neurodegenerative diseases from MRI scans is crucial for diagnosis but challenging due to complex multivariate patterns.
• Deep learning models excel at feature extraction but require vast datasets, which are often unavailable for medical imaging tasks.
• Existing methods struggle with anatomical variability inherent in MRI scans.

Purpose of the Study:

• To develop a parameter-free deep learning-like method for MRI-based neurodegenerative disease classification.
• To overcome the data limitations typically associated with deep learning in medical imaging.
• To improve the accuracy and interpretability of MRI classification for neurological disorders.

Main Methods:

• Utilized a three-dimensional scattering transform, a non-learnable deep convolutional neural network analogue.
• Applied the scattering transform to linearize anatomical variations in MRI data, enhancing disease pattern separability.
• Tested the method on brain morphometry in Alzheimer's disease and white matter damage in HIV patients.

Main Results:

• The scattering transform effectively captured essential features for disease classification from MRI scans.
• Achieved high accuracy in semi-supervised learning for distinguishing progressive versus stable Mild Cognitive Impairment (MCI), reaching 82.7%.
• Demonstrated effectiveness in classifying Alzheimer's disease and HIV-related white matter damage.

Conclusions:

• The scattering transform offers a powerful, data-efficient alternative to deep learning for neurodegenerative disease classification using MRI.
• This method enhances the separability of disease states by addressing anatomical variability.
• The approach provides a valuable tool for medical image analysis, with potential for improved diagnostic accuracy and interpretability.