Hizlari gonbidatua: Alexander von Lühmann. Recent advances in learning from fNIRS data for optical brain imaging and decoding in the everyday world.

What: Recent advances in data-driven approaches for fNIRS-based optical brain imaging

Where: BCBL Auditorium and Auditorium zoom room (If you would like to attend to this meeting reserve at info@bcbl.eu)

Who: Alexander von Lühmann, PhD. Senior Research Lead, Technical University of Berlin & BIFOLD, Berlin, Germany.

When:  Friday,  May 29th at 12:00 PM noon.

Advances in wearable functional near-infrared spectroscopy (fNIRS) and diffuse optical tomography (DOT) are rapidly expanding opportunities to complement fMRI by studying the brain in ecologically valid environments. Transitioning from well-controlled laboratory settings to the dynamic, complex, and multisensory environments of the everyday world presents a range of significant challenges, particularly in signal acquisition and processing. However, the increasing ease of acquiring larger data also paves the way for new, powerful data-driven approaches that can be used as a remedy. Research at the Intelligent Biomedical Sensing (IBS) Lab operationalizes this shift through the Cedalion toolbox. Cedalion provides a unified framework for reproducible signal analysis across heterogeneous datasets, multimodal integration, DOT reconstruction, and development of machine learning–driven methods.

In my talk, I will highlight some recent results from our lab that use machine learning to improve generalization and performance in fNIRS-based brain imaging and decoding: I) MRI-free data-driven head modeling using photogrammetry enables substantially improved DOT localization compared to atlas models, providing a practical alternative to subject-specific MRI for real-world applications. II) Systematic combination of high-density DOT, short-separation regression, and parcel-space features yield robust gains in single-trial decoding accuracy and cross-dataset generalization. III) Cross-modal learning in a shared parcel space enables transfer of large-scale fMRI priors to DOT, improving decoding performance and mitigating data scarcity. IV) Self-supervised spatio-temporal graph learning on parcel-space DOT data enable the extraction of transferable, network-level representations that could improve generalization across subjects and tasks, towards an fNIRS/DOT foundation model.