Identification of rare cortical folding patterns using unsupervised deep learning.

Area of Science:

• Neuroscience
• Medical Imaging
• Artificial Intelligence

Background:

• Cortical folding patterns are unique to individuals but follow general species-specific organization.
• Rare folding patterns are challenging to identify due to high inter-individual variability, yet some are linked to neurodevelopmental disorders.
• Identifying rare folding patterns could lead to novel biomarkers for diagnosing and understanding brain development.

Purpose of the Study:

• To develop and validate a novel unsupervised deep learning approach for identifying rare cortical folding patterns.
• To assess the method's ability to detect deviations in folding morphology.
• To evaluate the generalizability and clinical relevance of the proposed technique.

Main Methods:

• Preprocessing brain MR images to focus on folding morphology.
• Training a beta variational auto-encoder (β-VAE) on inter-individual folding variability to identify outliers.
• Comparing the detection power of the latent space and reconstruction errors using synthetic data and real-world cases.

Main Results:

• The β-VAE effectively encodes relevant folding characteristics, interpretable through its generative power.
• Latent space and reconstruction errors provide complementary information for identifying diverse rare patterns.
• The method demonstrates generalization across different brain regions and datasets, showing promise in epilepsy patients.

Conclusions:

• The proposed deep learning method successfully identifies rare cortical folding patterns and quantifies deviations.
• This approach holds potential for discovering novel biomarkers for neurodevelopmental disorders and other neurological conditions.
• The technique shows promise for clinical applications, particularly in identifying patterns in patients with drug-resistant epilepsy.