The application of the Hopfield neural network for the multispectral unsupervised classification of MR images is reported. Winner-take-all neurons were used to obtain a crisp classification map using proton density-weighted and T(2)-weighted images in the head. The preliminary studies indicate that the number of iterations needed to reach ;good' solutions was nearly constant with the number of clusters chosen for the problem.

Magnetic resonance angiography has matured to the point where clinically useful images can be acquired in half an hour or less. In this paper, the role of 3D imaging techniques is primarily considered. Specifically, the optimal imaging parameters, sequences, and reconstruction techniques are evaluated for moving spins. A variant of FISP known as ROAST with low flip angles, short repeat times, and a thick slab has been found to yield the best 3D survey scan of the cranial vessels with roughly 1 X 1 X 1-mm3 resolution in each of the processed images (slices). For the faster flowing carotids, a sagittal scout with as short a TE as possible is required to avoid spin dephasing. Localization is accomplished in both cases by acquiring thin slab 3D, thin partition, larger flip angle, longer repeat time FLASH sequences. Different choices of dephase/rephase sequences and directions are also reviewed. These choices are discussed from a practical and theoretical perspective. In particular, improvements in contrast and resolution are evaluated using half-Fourier, 512 acquisition, small fields of view and constrained reconstruction for both rephased gradient echo sequences and dephased thin slice long TR spin-echo sequences. A resolution of 0.5-0.75 mm is recommended to obtain sufficient image quality for consistent clinical interpretation of stenoses and vessel abnormalities.

No abstract