Application of a fuzzy neural network to medical imaging segmentation (LGN.4503)

The definition of the left ventricular (LV) endo and epicardial contours within medical images of cardiac patients is a necessary condition for the analysis of the global and regional functionality of this chamber. In daily clinical practice, imaging modalities, such as Magnetic Resonance Imaging (MRI) which produce three-dimensional (3D) data sets (including the LV), require the interpretation of tens to hundreds of slices. It has been widely accepted that the visual interpretation of the images by a specialist is very subjective. In addition, the manual tracing is tedious, time-consuming and also hampered by significant inter- and intra-observer variabilities. As a result, there is a great need for automated segmentation techniques, by which a computer defines the outlines (contours) of the myocardium with a high degree of accuracy and precision.

In this grant proposal an automated segmentation procedure is described, which combines multi- resolutional techniques and a fuzzy neural network, in order to classify the left ventricular boundaries (contours) in magnetic resonance images. By a multi-resolution approach an image (or an image series) can be composed at different resolution levels after which the image elements and their characteristic features are fed to the input of a fuzzy neural network. During a training phase, this network can learn and analyze the image features. After the training, a network will be generated with a minimum number of neurons and inter-layer connections. A trained network can carry out the classification of anatomical structures in terms of fuzzy rules such as ''If Feature Ai is Large then image element belongs to structure Bj''. User-defined fuzzy rules can also be added to the network. A pilot study carried out in our laboratory has shown excellent results obtained by such a network. Experience and results obtained in our laboratory so far with the application of a multi-resolutional segmentation technique and a neural network, have indicated the reliability of these techniques.

This project will be carried out in three phases. Firstly (Phase 1), after a general introduction into the cardiovascular research literature, into fuzzy neural networks and into the LKEB software infrastructure, about 200 MR image series of different subjects will be collected by the post-doc. In the second Phase a number of multiresolution techniques , e.g. pyramidal and wavelet-based, will be implemented to generate the multi-resolution image representations. In addition, image element values of these multi- resolution representations will be coded as vectors which will be fed into a fuzzy neural network. Also in the same phase, manually delineated outlines of the LVs, drawn by a number of specialists, will be used to define the output of the network. The second phase of the project also includes the feature analysis, by which the inputs of the network (image element features) will be analyzed by the fuzzy neural network; those superfluous inputs which do not discriminate between the LV and the background will be removed. The trained network will be evaluated by a test set (image series) and if necessary a user- defined rule set will be added to the network. In an optimization phase the network will be retrained and reevaluated. Finally, in the third Phase of this project the segmented approach will be integrated in the Magnetic resonance Analysis Software System (MASS), In addition, based on a fault analysis of the upgraded MASS , the trained network will be optimized again. As a result of a general approach of image segmentation in this project, it will be generally applicable as a solution method to solve related problems in other (cardio)vascular imaging modalities, such as intravascular ultrasound (IVUS) and X- ray angiography.

WWW Links

• Laboratory for Clinical and Experimental Image Processing

Er is één bedrijf bij dit project betrokken.

Prof.dr.ir. J.H.C. Reiber
AZ Leiden
Lab.voor Klinische en Exp. Beeldverw.
Afdeling Radiodiagnostiek
Gebouw 1C3Q-50
Postbus 9600
2300 RC LEIDEN.