Fuzzy Neural Networks

Application of a fuzzy neural network to medical image segmentation

Dutch Technology Foundation, STW: LGN 4508
Faiza Behloul, Ph.D


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.
Much work has been described on automated LV contour detection in cardiac MR images. Techniques have been described based on, among others, active contours and balloons, dynamic programming or pixel/region classification. In general, two major problems hamper many of these contour detection methods:

  1. An expert drawn contour does not always coincide with the location of the strongest local image features. For instance, many experts exclude the papillaries by drawing the LV ENDO border as a convex hull around the blood pool, somewhat 'outside' of the strongest edge. Therefore a contour detection method should not only be sensitive to the strongest local image evidence, but it should be adaptive to new examples and observer preference.
  2. Due to noise and image artifacts in routinely acquired clinical images, a-priori knowledge about the shape and image appearance of the LV is essential to achieve robust localization performance.

In this project, we aim to develop an automated segmentation procedure is adaptive to observer preference, by combining multi-resolution techniques and fuzzy neural networks, in order to classify the left ventricular boundaries (contours) in magnetic resonance images.


To develop a robust and adaptive segmentation method for cardiac MR and CT images based on multi-resolution and fuzzy neural network concepts.


In this project, we have developed a multi-resolution region classification method based on fuzzy clustering. In addition, we have developed an optimal method to automatically configure radial basis function networks (= fuzzy inference systems under specific conditions). Finally, we have developed a method to incorporate a-priori shape knowledge and appearance knowledge by simulating a virtual autonomous mobile robot that delineates the organ. A fuzzy neural network is trained on a representative set of examples by combining local image information and global a-priori shape knowledge. It is applied to provide the vehicle with recognition skills to delineate and label myocardial segments/regions it is passing through. The neuro-fuzzy techniques make the approach adaptive to observer preference, and robust with respect to noise and image artifacts. This latter algorithm has shown high robustness in the segmentation of cardiac MR and CT images, and has been integrated in the MASS CT software, which is distributed through Medis medical imaging systems.

Moving Turtle



The project was finalized in July 2001.


For further information, please contact:
B.P.F. Lelieveldt PhD,
Division of Image Processing
Department of Radiology, 1-C2S
Leiden University Medical Center
P.O. Box 9600
2300 RC Leiden
The Netherlands
Tel. +31 (0)71 526 2285
Fax. +31 (0)71 526 6801
e-mail: B.P.F.Lelieveldt@lumc.nl