Just outside the visible part of the brain are many low-intensity-but-not-zero voxels. These are caused by Fourier ringing and ghosting artifacts in the MRI acquisition and reconstruction process. Normally, you don't see these, since their intensity is maybe 10% of the brightness of the main part of the image.

However, you are dividing by the mean 'b'. Where the image has low-but-not-zero intensity, b is small and so division by b will scale things up relative to the inside of the brain. As a result, you see the weird artifacts being amplified to the same scale as the brain. And some of these b values may be tiny, so you may get some huge voxels outside the brain, after scaling.

Then, since your input image is shorts, AFNI will rescale the output to be shorts by default. This leads to a scaling artifact, since shorts only store 16 bits of value, but AFNI has to allow for the huge values scaled up by division by small b values. As a result, the normal sized values end up getting squashed to a low number of bits, and you see the quantization effect in your 'bad' time series.

The solution you have is a good one. However, if you don't want to suppress the small-but-nonzero voxels, you can get around the issue by using the '-datum float' option to 3dcalc. This will store the output dataset in float format instead of short format, and thus avoid the scaling/quantization artifact.

By the way, nice use of image attachments -- this is why we enabled them!