We describe a general method for extraction of an image object based on an approximate knowledge of its shape and its position in the image. The method is a combination of the watershed approach and neural networks. It is applied to high field MR images of a section of a spinal cord for the extraction of the shape of the grey matter.

This paper presents a novel medical image compression technique for inhomogeneous spatial reconstruction of MR images of the brain. The images are decomposed to various scales using the wavelet transform and a new multiscale segmentation algorithm is used to select areas of high diagnostic importance (regions of interest-ROI). Those areas, corresponding to brain tissue, tumours, and other structures in the head are compressed for maximum reconstruction quality, while neighbouring areas are coarsely approximated. The background is rejected since it contains only noise and no useful data. The quality of the reconstructed image is very good, even at low bit-rates, since the bit allocation is performed after diagnosis and reflects the diagnostic importance of each region. No useful data is lost at the selected ROI.

Although a considerable amount of research has been devoted to the registration of medical images, a good proportion of this has been performed on rigid regions of the anatomy. This has tended to limit the practical value of such research. The registration of medical images which are non-rigid in nature is a seemingly difficult and not fully investigated problem. This paper presents an elastic transformation method which is used to perform matching of non-rigid medical images, that is, two-dimensional images of a region of the body such as the thorax or abdomen which exhibits a high level of elasticity. Some forms of elastic transformation may be more localized in scope than others, which leads to the categorization of elastic transformation methods into local and semi-local methods. This paper demonstrates the use of local elastic spatial transformations in matching non-rigid medical images. These transformations are illustrated by matching corresponding computed tomography and magnetic resonance images of the thorax.

It is clear that statistical classification of MRS has considerable potential as an accurate diagnostic advice tool. However, sample sizes are not yet sufficient to guarantee performance in large-scale clinical trials. In addition, the most commonly used classification strategies, namely Principal Components Analysis of in vitro spectra and Linear Discriminants Analysis of in vivo data, while offering considerable accuracy, are not optimal even within the currently available linear statistical classifiers. The issue of non-linearity and the consequent need of neural network analysis remains rests also on the outcome of larger studies, although the performance of these methods has already been shown to be competitive with that of statistical methods. Further work on principled methods for neural network design and rule-extraction offer the potential, in the near future, for high-performing, transparent non-linear classifiers. Alternative algorithms can also be used for automatic labelling, of clustering, of spectra. Overall, the results of the preliminary studies reported here are encouraging both for the accuracy that has been achieved and the consistency of the variable relevance ranking with clinical expectation. (9 pages)

An automated data analysis approach has been developed for processing of ¹H NMR from tumour extracts. At present, the only manual interventions involve spectrum phasing (although automation methods are available) and choice of the number of principal components (usually chosen to account for about 99% of data variance). Unsupervised learning was important for identifying errors in the automatic processing scheme and for finding outliers (e.g. due to technical failures during extraction or NMR spectroscopy). it can also reveal underlying structure in the dataset (i.e. which classes of samples may be most easily separated). Factor analysis was useful for reducing data dimensionality (important for subsequent analysis) and, after vector rotation, for identifying important biochemical metabolites. Supervised learning (backpropagation NN) provided a robust classification method and was good for distinguishing between meningiomas and other types of brain tumour. Genetic programming analysis of a subset of these data gave comparable classification to NN but with quite simple `programs', facilitating biochemical interpretation. In general, classification of astrocytic tumours according to grade was not reliable based on ¹H NMR spectra from chemical extracts. Although in vivo ¹H NMR spectra of higher grade brain tumours might be characterised by elevated lipid signals, this information will be lost during extraction of water-soluble metabolites (as here). It would be expected that the best classification based on in vivo ¹H NMR spectra would involve short echo-time measurements since these are most sensitive to glutamine signals (important for distinguishing tumour type) as well as lipid signals (possibly dependant on tumour grade). (3 pages)