Contrastive Self-supervised Learning for Neurodegenerative Disorder Classification

Area of Science:

• Neuroimaging
• Artificial Intelligence
• Neurology

Background:

• Neurodegenerative diseases like Alzheimer's disease (AD) and frontotemporal lobar degeneration (FTLD) cause distinct brain volume loss detectable via MRI.
• Supervised machine learning for disease classification requires extensive, expertly labeled datasets, which are often difficult to obtain.
• Self-supervised learning (SSL) presents a promising alternative for training models without requiring labeled data.

Purpose of the Study:

• To investigate the application of SSL models for distinguishing between different neurodegenerative disorders using T1-weighted MRI scans.
• To assess the interpretability of SSL models in identifying disease-specific brain atrophy patterns.
• To evaluate the performance of SSL models compared to state-of-the-art supervised methods.

Main Methods:

• A deep convolutional neural network was trained using contrastive self-supervised learning as a feature extractor.
• A single-layer perceptron served as the classification head for downstream tasks.
• The model was trained and validated on 2694 T1-weighted MRI scans from ADNI, AIBL, and FTLDNI cohorts, including cognitively normal controls, AD, and FTLD subtypes.

Main Results:

• The SSL-trained feature extractor demonstrated generalizable and robust representations for classification.
• The model achieved 82% balanced accuracy for AD vs. cognitively normal (CN) on test and 80% on holdout datasets.
• For behavioral variant frontotemporal dementia (BV) vs. CN, the model attained 88% balanced accuracy.
• Integrated Gradient analysis highlighted hallmark atrophy regions: temporal gray matter for AD and insular for BV.

Conclusions:

• SSL methodology can effectively utilize unannotated neuroimaging datasets for training robust and interpretable machine learning models.
• The developed SSL models perform comparably to supervised deep learning approaches for neurodegenerative disease classification.
• SSL offers a viable strategy for leveraging large, unlabeled neuroimaging data in the study of neurological disorders.