Proceedings Article | 13 March 2010
KEYWORDS: Image segmentation, Skin, Medical imaging, Image processing algorithms and systems, Tumors, Mammography, Tongue, Magnetic resonance imaging, Image processing, Distance measurement
Object with blurry boundary is a very common problem across image modalities and applications in medical field. Examples
include skin lesion segmentation, tumor delineation in mammogram, tongue tracing in MR images, etc. To address
blurry boundary problem, region-based active contour methods have been developed which utilize global image feature
to address the problem of fuzzy edge. Image feature, such as texture, intensity histograms, or structure tensors, have also
been studied for region-based models. On the other hand, trained domain experts have been much more effective in performing
such tasks than computer algorithms that are based on a set of carefully selected, sophisticated image features. In
this paper, we present a novel method that employs a learning strategy to guide active contour algorithm for delineating
blurry objects in the imagery. Our method consists of two steps. First, using gold-standard examples, we derive statistical
descriptions of the object boundary. Second, in the segmentation process, the statistical description is reinforced to achieve
desired delineation. Experiments were conducted using both synthetic images and the skin lesion images. Our synthetic
images were created with 2D Gaussian function, which closely resembles objects with blurry boundary. The robustness of
our method with respect to the initialization is evaluated. Using different initial curves, similar results were achieved consistently.
In experiments with skin lesion images, the outcome matches the contour in reference image, which are prepared
by human experts. In summary, our experiments using both synthetic images and skin lesion images demonstrated great
segmentation accuracy and robustness.
