We propose a modification of Wells et al, technique for bias field est
imation and segmentation of magnetic resonance (MR) images, We show th
at replacing the class other, which includes all tissue not modeled ex
plicitly by Gaussians with small variance, by a uniform probability de
nsity, and amending the expectation-maximization (EM) algorithm approp
riately, gives significantly better results. We next consider the esti
mation and filtering of high-frequency information in MR images, compr
ising noise, intertissue boundaries, and within tissue microstructures
, We conclude that post filtering is preferable to the prefiltering th
at has been proposed previously, We observe that the performance of an
y segmentation algorithm, in particular that of Wells et al, (and our
refinements of it) is affected substantially by the number and selecti
on of the tissue classes that are modeled explicitly, the correspondin
g defining parameters and, critically, the spatial distribution of tis
sues in the image, We present an initial exploration to choose automat
ically the number of classes and the associated parameters that give t
he best output, This requires us to define what is meant by ''best out
put'' and for this we propose the application of minimum entropy, The
methods developed have been implemented and are illustrated throughout
on simulated and real data (brain and breast MR).