Sign in

In this lecture a multi-region segmentation method based on fuzzy intensity-connectedness is described. The first proposed seed-segmentation method is the original work of Seeded-Region-Growing (SRG) introduced by Adams and Bishop [1]. Despite its simplicity, SRG results are very good, but it was often pointed out that they are dependent on the order of analysis. To this end, some solutions have been proposed [2]. In all these works no fuzzy measure neither fuzzy processing are applied. The first segmentation algorithms based on fuzzy connectedness were independently proposed by Dellepiane et al., introducing in [4] the intensity-connectedness concept, and by Udupa et al. who defined the local fuzzy relation called "affinity" in [5]. In the present lecture, the adaptive growing mechanism, originally proposed in [3,4] for fuzzy intensity-connectedness measurement, is extended to the multi-seed case, to the volumetric third dimension, and to the multi-temporal analysis. Such a growing mechanism, starting from a small number of seeds, is adaptive to the actual data content and is able to correctly take into account local and global connectedness relationships. It can be proved that this growing mechanism assures the best path selection and turns to be completely independent of the order of analysis. This allows to be sure that the maximum membership decision step required for the extraction of fuzzy connectedness is not affected by any polarization or error. Such an aspect is mainly critical when analyzing real images, where contrast is poor and blurred, and spatial inhomogeneities are present, in addition to noise. The obtained result is a fuzzy segmentation where a membership value is associated to each analyzed spel (i.e., a pixel in the 2D space or a voxel in the 3D space, respectively) related to each seed-class. The de-fuzzification gives rise to a hard result, by applying the maximum membership criterion as usually applied in fuzzy cluster analysis. At the same time, the uncertainty degree associated with the classes remains available for each analyzed spel. In such a way it is possible to discard doubtful assignments, when one should prefer not to decide instead of taking a wrong decision. The application of the method allows to focus the attention to some regions of interest (ROI), avoiding the global segmentation of the whole image when it is not required. The handling of uncertainty allows to avoid the use of any parameter. Even though the method can be applied to any kind of digital image or digital volume, performance evaluation is presented in the biomedical domain, referring to the results obtained from two standard image databases from MRI, the former made of synthetic brain volumes, the latter made of real brain volumes. A quantitative performance evaluation is carried out on the two data sets where sensitivity, specificity, and accuracy for the segmentation of white matter, gray matter and cerebrospinal fluid are computed. References 1. Adams, R., Bischof, L.: Seeded Region Growing. In: IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 16-6, pp. 641-647 (1994) 2. Mehnert, A., Jackway, P.: An Improved Seeded Region Growing Algorithm. In: Pattern Recognition Letters, vol.18, pp.106-1071 (1997) 3. Dellepiane, S., Fontana, F., Vernazza, G.: Nonlinear Image Labelling for Multivalued Segmentation. In: IEEE Trans. on Image Processing, vol. 3, pp. 429-446 (1996) 4. Dellepiane, S., Fontana, F.: Extraction of Intensity Connectedness for Image Processing. In: Pattern Recognition Letters, vol.16, pp.313-324 (1995) 5. Udupa, J.K, Samarasekera, S.: Fuzzy Connectedness and Object Definition: Theory, Algorithms, and Applications in Image Segmentation. In Graphical Models and Image Processing, vol. 58, pp. 246-261, (1996).