Contrastive Learning for Neural Fingerprinting from Limited Neuroimaging Data

A significant challenge currently confronts neuroscientific research: the scarcity of labelled neuroimaging data, which restricts the capabilities of supervised deep learning models. To tackle this, we have developed three approaches to amplify the effectiveness of deep learning in neuroimaging, particularly when dealing with limited data. Our research delves into neural fingerprinting, a method that distinguishes individuals based on unique brain activity patterns, proving robust even in scenarios with limited data. We also introduce a flexible data augmentation technique suitable for a wide range of neuroimaging modalities. This method is specifically designed to enhance model training, even in the face of small datasets. Furthermore, we investigate the incorporation of contrastive learning into neural fingerprinting. This integration aims to improve the model's adaptability and performance, especially when encountering new, unseen data. Through these strategies, we have achieved an accuracy of approximately 98% in identifying individuals from various functional connectivity profiles. The contrastive learning approach, in particular, has shown significant versatility, especially in scenarios involving subjects not seen during the training phase of the model. This study contributes significantly to the field of neuroimaging, enhancing the accuracy and applicability of neural fingerprinting and laying the groundwork for future research in diverse neuroimaging contexts and methodologies. Our work opens new avenues for personalised medicine and advanced brain research, addressing the challenges of limited data.