ICHOR: A Robust Representation Learning Approach for ASL CBF Maps with Self-Supervised Masked Autoencoders

A multi-institutional research team led by Xavier Beltran-Urbano from the University of Pennsylvania has introduced ICHOR, a breakthrough self-supervised learning framework specifically designed for analyzing Arterial Spin Labeling (ASL) cerebral blood flow maps. ASL MRI allows non-invasive measurement of brain perfusion without contrast agents, making it ideal for repeated studies, but progress in AI applications has been hampered by variable image quality across different scanners, protocols, and limited labeled datasets. ICHOR addresses these challenges by using masked image modeling with a Vision Transformer backbone to learn robust representations from unlabeled data, creating what the researchers describe as a "general-purpose encoder" for various downstream neurological analysis tasks.

The technical breakthrough centers on ICHOR's training on what the team calls "one of the largest ASL datasets to date"—comprising 11,405 ASL CBF scans from 14 different studies across multiple sites and acquisition protocols. This scale enables the model to learn transferable features that generalize across the heterogeneity of real-world medical imaging data. In evaluations, ICHOR outperformed existing neuroimaging self-supervised methods adapted to ASL on three diagnostic classification tasks and one regression task predicting ASL map quality. The researchers will release pre-trained weights and code publicly, potentially accelerating AI development in neuroimaging by providing a powerful foundation model that researchers can fine-tune for specific clinical applications without starting from scratch.