Kevin A. Lee
Breast cancer is the second leading cause of cancer mortality in women. Mammography remains the best method for early detection of cancers of the breast, capable of detecting small lumps up to two years before they grow large enough to be palpable on physical examination. X-ray images of the breast must be carefully evaluated to identify early signs of cancerous growth. Segmenting, or partitioning, radiographic images into regions of similar texture is often performed during the process of image analysis and interpretation. The relative lack of structure definition in mammographic images and the subtle transition from one texture to another makes segmentation extremely difficult. The task of classifying different texture regions can be considered a form of exploratory analysis, since a priori knowledge about the number of different regions in the image is generally not known. This paper presents a preliminary examination of an image segmentation technique based on the Kohonen Self- Organizing Feature Map (SOM). The SOM network lends itself well to this problem for two reasons. First, such a network can be trained to recognize and classify regions exhibiting similar internal structure. It learns in an unsupervised mode, requiring no a priori knowledge about the number or nature of regions to be classified. Another important feature of the SOM is its topology-preserving behavior. The competitive learning algorithm employed by the network ensures that regions close together in the input space will maintain their relative proximity in the output space. This order-preservlng characteristic of the SOM makes makes it a good candidate for spatially-oriented problems such as image segmentation. The choice of node number for the competitive layer determines the maximum number or classes into which image regions can be partitioned. This paper presents a method of region classification using a simple SOM network and explores the effect varying the number of neurons in the competitive layer has on the resulting segmented image.