Interactive Kohonen neural network applications in medical image processing
|School||Qufu Normal University|
|Keywords||Kohonen neural network Image segmentation Output node Interactive PDS|
The segmentation of the most important application is medical image segmentation. A lot of biomedical information are shown in graphic form, such as X-ray image, CT image and the ultrasound image, it makes the human visual from the surface to the internal extension. People can use them to obtain useful information on the anatomy, biochemistry and physiological functions of internal organs. Lesions in medical images are sometimes in gray scale, shape similarity with the surrounding normal tissue, are not easily distinguishable with the naked eye, so the need for image segmentation, the lesions showing obvious. Self-organizing neural network is an instructor unsupervised learning algorithms by learning a set of data it can extract the important features or some inherent regularity (such as the distribution of characteristics, or some kind of feature clustering). This learning algorithm, based only on the properties of the input data and adjusting the weight, and then completed to the learning environment, tasks such as automatic classification and clustering. This article selection Kohonen neural network for medical image segmentation. Kohonen neural network was trained to determine the number of clusters (ie, the number of output nodes), usually there are two: (1) is proposed in accordance with the Hunts-beiger et al, to determine the number of clusters 4; (2) in accordance with the peak number of points in the image histogram to determine the number of clusters. The two common methods, in some applications, be able to obtain a more satisfactory training results. However, in medical image segmentation, it can not achieve the desired results. Because, for an image, its histogram is fixed, and thus the peak number of points is fixed, for image segmentation result is unique. So, whether the provisions of the number of clusters, or regulations in accordance with the peak number of points to determine the number of clusters are two ways to make the segmentation of medical images become very mechanical, lost adaptability, thereby affecting analysis of the doctors of the disease. Flexible segmentation of medical images, we propose interactive Kohonen neural network image segmentation. For a medical image, the number of clusters to determine a predetermined pattern, which means that the number of nodes of the Kohonen neural network output layer prior uncertainty, but by the doctor according to the analysis of needs and conditions decide for themselves. The number of clusters of images from a doctor entered into the computer, the neural network to determine the number of output nodes, the network training. This established a man-machine interactive neural network training. Take advantage of this interactive network training with a medical image, the doctor can be interactive Kohonen neural network applications in medical image processing to transform the number of clusters flexible to different medical image segmentation results, and from pieces of image segmentation, image segmentation elect to contribute to disease analysis. Kohonen neural network for image segmentation, the computation is very large. Improve processing speed in order to reduce the amount of computation, we introduce a tower-type data structure (P DS), the image is divided into different levels of quantitative image, making the premise does not affect the accuracy of image segmentation, computation is greatly reduced, thereby enhancing the the speed of image segmentation.