Segmentation of ultrasonic images using learning vector quantization network

In this study, a novel method for biomedical image segmentation that provides an effective classification of similar pixel groups and their following extraction into separate regions is proposed and analyzed. Original image pixels intensities are used as a source of input data. The classification is carried out by means of a learning vector quantization network, which is employed to obtain the extraction of main classes (structures, tissues, artifacts etc.) presented in the image. Since this type of neural network supposes that number of classes is known a priori and learning algorithm requires supervision, a participation of operator is necessary to create the set of input data. The results obtained from ultrasonic test image processing show the feasibility of the method for similar classification problems. A quantitative evaluation for the segmentation results was conducted considering a comparison with a human assisted segmentation.