Wednesday October 14 2020, 10:00-11:00 AM EDT
The video recording of the session is below.
Use the commenting tool below to discuss the presentations given in this session. Please register as a member to be able to ask and answer questions https://fnirs.org/account/. You can register for free for this web content.
University of Pittsburgh, USA
Recent additions to the NIRS-Toolbox
University of Essex, UK
|Computational intelligence in fNIRS data analysis||
University of Houston, USA
|Assessment of fNIRS data quality at individual and group levels||
Northeastern University, Boston, USA
|Wearable and modular fNIRS probe||
Washington University, St. Louis, MO, USA
|Wearable high density diffuse optical tomography||
Washington University, St. Louis, USA
|Illumination childhood development with high density diffuse optical tomography||
Massachusetts General Hospital, USA
|Optimization of experimental and computational approaches for cerebral blood flow monitoring||
|WeP2||Panel Discussion||Software and Hardware Moderators Rickson Mesquita & Mari Franceschini||
Return to main chat page https://fnirs.org/conferences/fnirs-datablitz-2020-chat/
Ted, do your brain analyzer object classes integrate well with the SNIRF format? Seems that this will be super powerful if we can load the SNIRF file into your objects.
Yes, we allow import and export of SNIRF using the commands nirs.io.loadSNIRF and .saveSNIRF. In particular, importing via SNIRF will preserve all the 3D registration, any loaded subject demographics and metadata, and stimulus timing information
Love the multimodal integration of your package. What level of EEG analyses can be performed? (E.g., microstates, FFT)
So far, we have some basic methods including filtering, PCA /ICA for artifact removal, spectral analysis, ERP estimation, and image reconstruction. We have been focusing on multimodal integration methods, for example using canonical correlation of mixed modality signals and joint image reconstruction.
Hi Ted– Lots of great improvements. I’d love to hear more about the short channel integration with FC connectivity. How does it operate differently than task-based GLM analyses?
This is work in progress that we are working with Pradyumna Lanka and Heather Bortfeld to compare methods. What we are comparing is preprocessing methods (e.g. using SS data as a filter) verses partial correlation methods, grangers causality, etc. It looks like partial correlation is going to win, but we are still working on how to deal with motion artifacts and noisy SS data through robust statistical models.
Amazing talk Dr. Huppert. What would happen if the short channels included as regressors are noisy?, Would that affect the outcome?
Do you recommend to add also 1st and 2nd principal components as regressors? How adding this will affect the data?
For the GLM model, we found that using a PCA decomposition of the SS data and using the components as the regressors is better than using the data directly. We still use all the components, but using the PCA redistributes the noise. We also use robust statistical methods (iterative outlier rejection methods) which is essential for the issue of noisy SS data.
Very intuitive figures. Would it be possible to save the data quality metric back in to the SNIRF file so it could be easily used by all toolboxes? Perhaps using the custom metadata tags that SNIRF supports? BIDS has the concept of a bad channel, but that might be too binary for your rich metric.
Thanks, Rob! It is definitely possible to save a simpler, binary output of the assessment into SNIRF/BIDS to the extent that the formats allow (Sam has all details on the code, and he will be involved in our BIDS effort)
Thanks for your fantastic ongoing work, Lucca, looking forward to having the tools fully available to the community. Can you elaborate a bit on the difference between motion artefact and poor optical coupling (the red and black bands) on the chart mapping data quality?
Thanks a lot, Anna! I will pass along the slides to you for better visual reference, but in essence we use a combo of SCI and PSP measures to separate movement artifacts (red pixels, high SCI + low PSP) from poor coupling (black pixels, low SCI + low PSP), and we do so on very short signal windows to keep our temporal resolution high. Does this sufficiently answer your question?
Great talk, Javier. Do you think that using computational intelligence to both preprocess and analyze activation may lead to overfitting?
Thanks Rickson, when using CI to model neural data, there is always the risk of including residual variations (i.e. noise), but a good strategy is to include a lot of strong learning constraints to penalize the complexity of the model (e.g. regularization) and also in the intermediate learning steps (e.g dropouts). If the model is well controlled in this respect, it should converge to the desired modeling phenomena and ignore the redundant effects. However, I would say that to leverage the automated learning process, it’s suggested some pre-filtering of the fNIRS signal before the AI modelling.
I am interseted in your thoughts on full head probe arrangements on different size heads. We find that 3.0cm separation is optimal for large males, but not for average size females. See http://fmri.org/wp-content/uploads/2020/10/nirscap-scaled.jpg for a reference.
This is a good question and something we are still trying to figure out.
In fiber based systems and NIRx, the optodes are fixed to 10-20 coordinates such that optode distance scales with head size. The advantage of this is that the channels are over roughly the same anatomical areas in all subjects. This allows you to just stack and average the data across subjects. However, the disadvantage of this is that SNR scales logarithmically with optode spacing. Thus, if you wanted to test the hypothesis of male vs female brain activation (or its even worse for something like child development) this is difficult statistically because your noise is unbalanced across the groups. This is not a trivial statistical problem. On the other hand, if you use fixed distances (say 3cm irregardless of the head size), then you need to adjust for the activity moving across different channels. We are working on analysis to account for this later problem. Since noise is logarithmic with distance, it looks like keeping source-detector distance constant across subjects is probably the better solution from a statistical point of view
Hi, What detector do you use? If these are APDs, arent you worried about high voltage? do you use any galvanic isolation?
We use hamamatsu Photodiodes. I’ll look up partno.
The PD that we use is from Hamamatsu, S12158-01CT. I think this is the same as used by Gowerlabs (Rob?) Seems like this PD might play the same role as the Hamamatsu 5460 APD has done for many of the fiber systems. It’s a great ultralow noise PD, with 4 mmx4mm area.
Great development! How robust against hair is your new wearable system? It seems your probes are placed on the occipital hairy region.
Yes. Over the hair. That was a design constraint, has to work everywhere. Seems to work OK. Pretty similar to fibers.
Thank you for the presentation Qianqian, really impressive work.
I was curious about which kind of microcontroller are you using for the modules, and wether each module has its own controller.
Overt language production changes the pCO2 in the blood, which strongly affects brain hemodynamics. Did you measure pCO2?
Very exciting, Joe! I am interested in the power line issues that you mentioned: is it mostly 60 Hz component (and harmonics) that you see? I imagine that Faraday cage around the probe helps!
Well there are true EM interference stuff, and so that’s why we have the foil around. We are looking to reduce the wait size of that, but that helps. But also we found we can get rid of what looks like EM, but fixing the power supply and ground planes.
I think the 60 Hz is gone. One source remaining is the digital signal from the A/D readout.
Great stuff Joe – really interesting system design. Is there a paper on the way?
should w all move to heterodyne DCS?