Fully automated algorithm to detect MRI noise/artifact?

I want to detect a “noisy” or “artifact-impaired” MRI image – basically a way to classify the image as noisy or not. For clarity, a “good” image looks like the first one, whereas a “bad” image might look like one of the other two.

STW-0113-#My Trajectory_20210614_091831_Slice 6-004-S1-CC0.6932CDH-0007-#My Trajectory_10Ma208-_9_1_6__Slice 3-007-S2-CC0.7973DUKE-0007-#My Trajectory-1_20150909_164101_Slice 6-045-S1-CC0.5456

I have a hunch that a measure like Shannon Entropy might be able to distinguish these and I’m about to do that experiment on my datasets.

What I wanted to ask, is how do I find prior research on the topic of simple detection? I tried the obvious searches – MRI artifact detection / noise detection – and find most results are about motion artifact detection and removal. That is interesting, of course, but I haven’t found anything yet on simple detection of the kind of images shown above; I’m not attempting to “correct” for the noise, just detect it. Any thoughts on keywords to use?

1 Like


I am not certain of the capabilities / techniques involved, but a group at Kitware has been developing an open-source project the includes traditional and deep learning methods for MRI QC.

Perhaps it will help with code samples/methods to consider.



1 Like

We had a short paper last year about it. We have a demo deployment. I am not sure whether it is still fully functional.

You might want to take a look at signal to noise ratio (SNR) classic quality metric, links here.