Contrastive self-supervised learning for neurodegenerative disorder classification | JoVE Visualize

Area of Science:

Neuroimaging and Machine Learning
Computational Neuroscience
Radiology

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 neurodegenerative disease classification requires extensive labeled datasets, which are challenging to acquire.
Self-supervised learning (SSL) offers a promising alternative for training models on large, unlabeled neuroimaging datasets.

Purpose of the Study:

To investigate the efficacy of self-supervised learning (SSL) models in distinguishing between different neurodegenerative disorders using T1-weighted MRI scans.
To assess the interpretability of SSL models in identifying hallmark brain regions associated with specific neurodegenerative diseases.
To evaluate the performance of SSL against state-of-the-art supervised methods in neuroimaging-based disease classification.

Main Methods:

Developed a two-part method comprising a deep convolutional neural network feature extractor trained with contrastive loss and a downstream single-layer perceptron classifier.
Utilized a dataset of 2,694 T1-weighted MRI scans from four cohorts, including cognitively normal controls (CN), Alzheimer's disease (AD) cases, and frontotemporal lobar degeneration (FTLD) phenotypes.
Employed Integrated Gradients for feature attribution to visualize and interpret model predictions.

Main Results:

The SSL-trained feature extractor yielded generalizable and robust representations for classification tasks.
The model achieved 82% balanced accuracy for AD vs. CN classification on test and holdout datasets (80%).
The model achieved 88% balanced accuracy for Behavioral variant frontotemporal dementia (bvFTD) vs. CN classification, highlighting temporal and insular atrophy patterns.

Conclusions:

Self-supervised learning (SSL) provides a robust and interpretable method for classifying neurodegenerative diseases from MRI scans without requiring diagnostic labels.
SSL models demonstrate performance comparable to supervised methods, enabling the effective use of large, unannotated neuroimaging datasets.
The interpretability of SSL models aids in identifying disease-specific neuroanatomical changes, such as temporal gray matter atrophy in AD and insular atrophy in bvFTD.