NYU introduced Neuro-JEPA for analyzing MRI brain scans

Researchers from NYU Langone Health released Neuro-JEPA — a specialized 3D model on the Vision Transformer architecture designed for neuroimaging data processing, with the model being pre-trained on an array of 1.55 million brain MRI scans, the model's architecture is focused on multimodal analysis of medical images, during testing the system demonstrated high accuracy, showing excellent results in 47 different data analysis tasks, including pathology classification and brain structure segmentation, the development uses self-supervised learning principles to extract features from three-dimensional medical data, allowing the model to effectively work with anatomical structures without the need to label each individual image, the model's weights and documentation are available to researchers in open access on the Hugging Face platform, the implementation of such models reduces the time for preparing diagnostic tools, automating the primary processing of large volumes of clinical data, the use of 3D transformers in neuroimaging allows for consideration of spatial relationships between brain departments, which are often missed when analyzing 2D slices.