A. Zaim and J. Jankun (USA)
Segmentation, transrectal ultrasound (TRUS), prostate, self-organizing map (SOM), gray-level co-occurrence matrix (GLCM).
The presence of strong speckle noise and shadow artifacts in transrectal Ultrasound (TRUS) images prevents accurate extraction of prostate using classical segmentation techniques. Most modern segmentation techniques adopt model-based approach such as active contour and others that are considered semi-automatic because they require initial seeds or contours to be manually identified. In this paper, we propose framework for automatic segmentation of ultrasound prostate images using Kohonen neural network. A set of morphological transformations are first applied to remove speckle noise. A new technique is then developed to remove ultrasound specific speckles using region-based thresholding and utilizing feature-based measures of gray-level co occurrence matrix (GLCM). Kohonen clustering network is employed to identify prostate pixels taking spatial information as well as GLCM measures, namely contrast and entropy, to form its input vector. The clustered image is then processed to produce a fully connected prostate contour. A number of experiments comparing the extracted contours with manually-delineated contours validated the performance of our method.
Important Links:
Go Back
