DeepBrainPrint: A Novel Contrastive Framework for Brain MRI Re-Identification has been accepted at the "Medical Imaging with Deep Learning" conference (MIDL 2023). The paper is the first to propose a semi-self-supervised contrastive deep learning approach, which retrieves brain MRI scans of the same patient from a medical imaging datasets and outperforms previous methods such as InfoNCE, SoftTriple, SimCLR, and BarlowTwins. Read the pre-print here.
DeepBrainPrint lays the foundation for a brain fingerprinting methodology powered by a semi-self-supervised deep learning pipeline, which enables accurate and generalizable subject re-identification through brain MRIs.
DeepBrainPrint trains a neural network to map a brain MRI to a numeric fingerprint that characterizes the unique morphology of a human brain, enabling image-driven re-identification of subjects. The fingerprint is robust to subject-related variability (e.g., ageing, disease progression) and domain shift variability caused by different contrasts or acquisition scanners.
Experiments results demonstrate that DeepBrainPrint outperforms both state-of-the-art deep metric learning techniques and established brain fingerprinting methods in terms of performance, efficiency, and generalization capabilities.
Besides re-identification, the resulting model can serve multiple purposes, including searching for brain scans with similar structural characteristics, such as shape, lesions, or atrophy, even if they are from different subjects.
