The largest sampled spatial frequency is directly linked to the synthesized beam size (i.e. the interferometer spatial resolution). The pixel size must be at least 1/2 the synthesized beam size to ensure Nyquist sampling in the image plane. However, Nyquist sampling is enough only when dealing with linear process while deconvolution techniques are non-linear process. We thus recommend to pick a pixel size between 1/3 and 1/4 of the synthesized beam to ease the deconvolution. Unless you have good reasons, you should not choose too large an image size nor too small a pixel size at least because it would considerably slow down any deconvolution algorithm.