CLEAN sometimes runs into an instability when working on extended sources of fairly uniform surface brightness. It produces obviously wrong `CLEAN stripes' in your image. The Fourier transform of these stripes tends to be found in unsampled parts of the u,v) plane. This is a complicated way of saying that the visibility data do not constrain CLEAN's interpolation between the measured visibilities sufficiently well to prevent it putting in badly wrong values.
One change to the Clark CLEAN that helps suppress CLEAN stripes is the SDI CLEAN (Steer, Dewdney, and Ito 1984). In this variant, implemented in SDCLN, any point in the residual image greater than some factor times the maximum residual point is taken as a CLEAN component. The factor is less than one. Thus, when the residual image has become very smooth, this avoids introducing ripples into it (which occur when subtracting the beam from just one location) which lead to the stripes. This algorithm is quite successful, although the implementation in SDCLN is slow.
Note that the Högbom, Clark, and SDI versions of CLEAN only allow you to CLEAN the inner quarter of your image properly because the image and beam are the same dimensions. For example, consider images that are 256 pixels in size (one dimension will do), with a point source located at pixel 192. A beam image centred on that point source will extend only to pixel 64, so that the full image cannot be CLEANed.