“pixel width” is actually a misnomer, since pixels are just sample points in space and don’t have an extent. What’s meant here is the pixel spacing, i.e. the distance between two sample points.
That’s determined by your image acquisition. If you use a calibrated system, the acquisition software might write the correct pixel calibration into your image metadata, and ImageJ (more specifically, Bio-Formats) will be able to read them.
But I wouldn’t always trust the metadata read by ImageJ, since it can get wrong in several ways. So be sure you know the scaling at which you acquire images. If necessary, you can simply measure the image of a ruler and calibrate the image yourself, using Analyze > Set Scale….