Cedalion puts optical brain imaging in one Python framework

Seven cloud-run Jupyter notebooks offer a new way into optical brain imaging: no local setup, just a guided path through Cedalion, a Python framework developed by researchers at BIFOLD and TU Berlin. The tutorial brings the field’s analysis tools into one environment, with the stated aim of making wearable brain data easier to work with, share and reproduce.

The underlying measurements are designed for life beyond the scanner. Functional near-infrared spectroscopy, or fNIRS, tracks brain activity through changes in cortical blood flow; diffuse optical tomography, or DOT, extends the approach with more information about location and depth. Because fNIRS is wearable, researchers can use it during movement and outside the lab, and can combine it with EEG — electrical brain recordings — and physiological signals.

Cedalion connects the chain that used to be split across software environments: simulating how light moves through tissue, estimating sensor positions on the head, checking signal quality, correcting artifacts, modeling multimodal time series with general linear models, reconstructing DOT images and applying machine learning. Its open-source design builds on the MATLAB tools Homer2/3 and AtlasViewer, carrying their methodological experience into Python. The work is described in a tutorial published in Neurophotonics.

What changes in practice? A researcher can develop a method, compare it with existing ones and share a complete workflow in the same environment instead of asking colleagues to rebuild the analysis by hand. Cedalion’s support for the open SNIRF and BIDS data standards, cloud-executable notebooks and automatic links to source publications also gives other teams a clearer route to rerun the work. That matters for studies involving brain decoding, clinical monitoring or motor and cognitive tasks tracked with wearable sensors.

The optimism has boundaries. Some functions, including photon simulation and machine-learning training, can be computationally expensive, although the authors describe GPU, CPU and cloud options as ways to address that cost. Researchers still have to judge whether machine-learning methods are methodologically sound, and systematic benchmarking of community pipelines on standardized datasets remains unfinished. The seven notebooks make the first step easier; they do not remove the need to test what follows.