Deformable-model-based segmentation techniques can overcome some limitations of the traditional image processing techniques. Currently developed deformable models can cope with gaps and another irregularities in object boundaries. However, they present problems in noisy images. Our approach is able to segment objects in noisy images by defining a new energy function associated with image noise and avoiding the tendency of contour points to bunch up. The model is validated for vessel segmentation on mammograms.

Segmentation in noisy images is an important and difficult problem in pattern recognition. Edge detection is a crucial step in this process. Current subjective and objective methods for evaluation and comparison of segmentation techniques are inadequate or not applicable to edge detection techniques. A general framework for segmentation evaluation in noisy images is introduced after a brief review of previous work. Several measures based on similarity between true and result segmented images are defined. These measures are, then, combined in a unique criterion as a proposed global measure of performance. The results indicate that this global measure can be helpful in the evaluation and comparison of segmentation techniques applied to noisy images.