Brain Hemisphere Dissimilarity, a Self-Supervised Learning Approach for alpha-synucleinopathies prediction with FDG...

Area of Science:

• Neuroscience
• Medical Imaging
• Machine Learning

Background:

• Idiopathic REM sleep Behavior Disorder (iRBD) is a key indicator for alpha-synucleinopathies like Parkinson's disease (PD) and Dementia with Lewy bodies (DLB).
• Predicting conversion during the prodromal iRBD phase is crucial for disease management and clinical trials.
• Existing supervised deep learning methods are limited by scarce longitudinal data and conversion information.

Purpose of the Study:

• To develop a self-supervised learning strategy for analyzing brain PET scans in iRBD patients.
• To pre-train a convolutional network using a novel hemisphere dissimilarity loss (HDL) on limited data.
• To predict conversion to PD or DLB using baseline 18F-FDG PET scans.

Main Methods:

• Proposed a self-supervised learning approach using 18F-FDG PET scans from iRBD non-convertor subjects.
• Introduced a novel hemisphere dissimilarity loss (HDL) function, extending Barlow Twins loss.
• Pre-trained a convolutional network and fine-tuned it for conversion prediction tasks.

Main Results:

• The proposed HDL method outperformed variational autoencoders in generating predictive brain features.
• Successfully pre-trained a network without disease-specific information, enabling analysis on small datasets.
• Demonstrated the utility of HDL for predicting conversion to PD and DLB.

Conclusions:

• Self-supervised learning with HDL is a viable strategy for analyzing neuroimaging data in iRBD.
• This approach can aid in predicting neurodegenerative disease conversion from prodromal stages.
• The method offers a promising avenue for improving early diagnosis and patient stratification.