Figure 1: the Architecture of a Kohonen Network. Each Input Neuron Is Fully Connected with 4.1 Intensity-image

the neurons in the map, but only a few connections are illustrated. The weight vectors ~ w of the neurons in the map form the pivot of the network. Figure 2: a) The training image. b) The a priori segmentation of the training image. c) A test image. d) The required segmentation of the test image. Figure 3: a) The training image. b) The a priori segmentation of the training image. c) A test image. d) The automaticallly obtained segmentation of the test image, on the basis of Figure 4: a) The training image. b) The manually obtained segmentation of the training image. c) A test image from the same three-dimensional dataset as the training image. d) The acquired segmentation of the image in c. Using as features L(0), L(1), L(2), L(4), L(8) and L(16). e) An image from a diierent dataset as the one the training image was obtained from. f) The acquired segmentation of the image in e. We have applied our strategy to segment various MR images of the head. Figures 4a and b illustrate the training image and a manually segmented version of it. A network was trained with feature patterns derived from the training image, and its weight vectors were labeled according the a priori segmentation. With this network several images, resembling to the training image, were segmented. The resulting segmentations are illustrated when zeroth order information at various scales was taken into account ((gures 4c-f). From the results we may conclude that the representations of the objects in the images are robust and fairly insensitive to shape and position variations, because a number of images with signiicant diiering intensity distributions could be segmented satisfyingly with the same network. 5 Conclusion A rst conclusion of our study is that we have developed a system which is capable of segmenting similar images, once it has been trained to segment a representative one. Apparently scaled diierential invariant features provide a means of describing pixel properties , on which basis pixels can satisfyingly be classiied. The multiscale approach ensures that the characteristics of pixels are not restricted to local properties, but extend to environments of varying extent. Applying features at multiple scales yields robust descriptions of the objects in the images, by feature patterns which are prototypical for the object, and supports the acquisition of homogeneous segments. Inclusion of gradient information improves the classiication of pixels near boundaries …