Discovery of Coherent Noise Source

On August 9th, we were making very sensitive noise measurements during testing of an investigational 64-channel receive array.  We observed some artifacts in the noise covariance matrix when testing the array.

We originally attributed the noise to an issue in the device under test. However, after repeating the test while turning off all power to accessory devices in the scan suite (TV, Eyelink, Biopac, Webcam, Lights, and Medlux table lights) we found an effect related to the room lights (Fig. 1).   The noise covariance matrices in Fig. 1b show correlations between channels when the room lights were turned on  – indicative that there is a coherent noise source related to the room lights.

This artifact may not appear in reconstructed MRI data.  The reasons is that a typical reconstruction performs noise whitening, effectively weighting down the contribution from channels with large noise power. This results in SNR maps (Fig. 1a) and an SNR histogram (Fig. 1c) that vary only slightly between the two conditions of room lights being on or off.  So the noise is there, but the reconstruction algorithms may remove the effect.

Based on this testing, we believe the noise was introduced from a bulb that went bad.  On the evening of August 11th, we replaced all the large bulbs in the scanner room and repeated our tests (Fig. 2).  The SNR maps and noise covariance matrices now show identical behaviour, irrespective of the condition of the room lights being on or off.  So, that’s good.   We further analyzed the last year’s worth of image quality data, which is all stored on Flywheel, and determined the onset of the problem was sometime on August 2nd, 2021 and it was resolved when the light bulbs were changed on the August 11th, 2021.

What does this mean to you?

Please check data you acquired between August 2nd and 11th.  Our  tests don’t show a significant SNR variation in the reconstructed images, but some reconstruction methods may fail to suppress the higher noise. If you find unusable data, please contact Laima Baltusis and she will arrange for replacement scan time to reacquire the data.

What is CNI doing to stop this problem from occurring in the future?

A number of events that can cause coherent noise to appear on MRI data: degradation in the room shielding, a failure in any electrical accessories in the room, or a failure in the MRI scanner hardware itself.  We acquire QA data once a week, but in this case we didn’t check it right away. We have now implemented a simple software app to monitor the Flywheel QA project week by week and show the results.  The data are consistently at a high SNR for more than a year, so this seems to have been a single case.

Over time we will continue to refine the types of QA data that we acquire in order to be sensitive to a larger variety of issues, and we will be quicker to read the QA data.  We also have regularly scheduled GE service maintenance that checks the system for issues such as these as well as many more.

You are the best steward of your data.

Prof. Nolan Williams’ group is to be congratulated for independently discovering a change in data quality on some of their scans and reaching out to CNI. Their messages to us were important. While we are improving our QA processes, we cannot promise to detect all possible sources of bad data.  We urge everyone to regularly check their data as soon as possible after acquisition for any change in quality or behavior — you can help the entire CNI community by making an early discovery of a data problem.

Thanks, and stay safe,

The CNI Team