First Experience with the Kernel Flow Device
Posted By David Boas, July 23, 2021
We received the much-anticipated Kernel Flow device on Tuesday July 20, 2020. This photograph shows the device out of the box. It was already fully assembled and transported with the helmet on this stand. In a separate box was a pre-configured computer to control the device and a USB-C cable with power supply to connect the device to the computer. Really quite simple to connect and set up.
We followed the quick start guide which led us through launching the acquisition software and calibrating the device. The calibration is done with the helmet on the stand. As far as I can tell from the messages on the screen, this calibration is confirming the pulsing of each of the 52 lasers and their arrival times on each of the detectors. I recall that aligning laser pulses across multiple sources to within 100 pico-seconds was always a challenge when we were building our own time domain fNIRS system 15 years ago [Selb2006]. Their solution of calibrating this electronically on the cleverly designed stand appears to be highly effective.
Next, the quick start guide instructed us place the helmet on my lap and place struts in 4 specific locations, choosing the appropriate length struts depending on the head size of the subject. This was very easy to do and is another clever solution to deal with variable head sizes. You can see these 4 white struts in the photograph. One is in the frontal region, one on the top of the head, and the other two are smaller and to the left and right sides.
In the photograph you can also see the layout of the 52 modules where each module has a source in the middle and is surrounded by 6 detectors. As I’ll describe below, measurements are made between these sources and detectors within a module, but also across the modules. In the photograph you can also see 6 dry electrodes for simultaneous EEG measurements. You can see the USB-C cable running to the back of the head and joining the helmet next to a dial that tightens wires to conform the helmet to the subject’s head.
Before making the first measurement, I had to create an account and establish my lab on the Kernel Portal in the Cloud. Then other members of my lab could create accounts and I could join them with the lab’s account. We could then set up studies and invite participants to a given study. I registered as a participant and accepted to be part of our first study “Having Fun”.
Once we placed the helmet on my head, the acquisition software first had us “tune the lasers.” I suspect this process was adjusting the laser powers based on the scalp coupling we obtained through my hair. We were able to visualize in real-time the photon counts and “scalp coupling index” for each detector, which is a nice feature for making sure that signal quality is acceptable before proceeding with an experiment. As we are used to doing with our own fNIRS systems, we would shift the helmet side to side and front to back to “comb” the fiber tips through the hair to improve the optical signal quality. The fiber tips are spring loaded and felt very comfortable.
For our first experiment, we ran a finger tapping task that was already set up on the Kernel computer. Basically, the Kernel acquisition computer is set up to work with two monitors, one for the operator and one to face the participant. Built into the acquisition software are several different experiments that it can present to the participant on this second monitor while acquiring the fNIRS and EEG data in a time synchronized fashion. I’ll leave our analysis of the brain activation results for a future blog post, and just focus here on our initial observations about the signal.
After acquiring 5 minutes of data, the computer automatically started uploading the raw data file to the cloud. Once it was in the cloud, we could then download a quality control report as well as SNIRF files that we could then analyze with our own software. The quality control report shows that on my head we were able to get strong photon count rates generally above 1 million photons per recording interval (about 5 ms) at both 690 and 850 nm for the nearest neighbor measurements at 10 mm (see figure below). We could see the mean total counts versus distance and see the expected exponential decay with distance out to about a 50 mm source-detector-separation (SDS) before the mean signal reached the noise floor of about 500 photons per recording interval (see figure below). It also reports to you the mean signal time courses within a module (as well as between modules). We are now working on adapting Homer and AtlasViewer to be able to process these time domain fNIRS data sets from the Kernel device, and hope to report soon our progress.
One thing that gives me a little pause is the huge amount of data that this device generates. This is a great thing. It has 52 sources and 312 detectors. It generates measurements from 2134 different source-detector pairs at two different wavelengths. For each source-detector pair it generates a full temporal point spread function (TPSF) at a rate of about 7 Hz. You can export the SNIRF file with just the first 3 moments of the TPSF or with 10 different temporal gates. Thus, in total, we have 2134 source-detector pairs per wavelength X 2 wavelengths X 5 measurements per second X 3 moments, which gives us 64020 measurements per second. But we will work through this awesome amount of data and hopefully get imaging results very soon.
Overall, the system appears to work very well. It has been very intuitive for us to operate. Once I slowed down and read the email about setting up the lab account on the Kernel Portal, that whole process was very easy. At first, a colleague over tightened the cap on my head which made it a bit uncomfortable after about 15 minutes. But this was easily resolved by not over tightening the cap and subsequently could wear it for much longer periods. The cap is a little heavier than it looks, but one has to consider the phenomenal amount of technology in this helmet and it is remarkable that it doesn’t weigh 100x more as vastly fewer numbers of time domain fNIRS channels weighed only a few years ago. And related to this, the outside of the helmet does get warm, but the inside remains a comfortable temperature for the subject. Again, given the very advanced technology in the helmet, I am not surprised by this. It indicates that the helmet design is effective at heat conduction away from the detectors to help reduce the noise floor as much as possible.
More to come…
