A training set of MR images of normal and abnormal heads was used to derive a complete set of orthonormal basis functions which converged to headlike images more rapidly than Fourier basis functions. The new image representation was used to reconstruct MR images of other heads from a relatively small number of phase-encoded signal measurements. The training images also determined exactly w.h/.ch .t .t .t .t~- encoded signals should be measured to nnmmme nnage reconstruction error. These signals were non-uniformly scattered throughout k.space. Experiments showed that head images reconstructed with the new method had less truncation artifacts than conventional Fourier images, reconstructed from the same number of signals.