Automatic segmentation of ultrasound images with clustering

Stefano Rovetta

The set of tools presented here can be used to support automated or semi-automated analysis of visual information in ultrasound images. The target application is identifying the region of interest in carotid images for automatic measurement of Intima-Media thickness. The tools are centered around partitional methods for clustering based on centroids. In this application, centroids are used as prototypes of parts of images and provide a segmentation of the images.

The specific clustering methods used belong to the "fuzzy clustering" family and have been developed by the Machine Learning and Soft Computing group, Department of Computer and Information Sciences, University of Genova, Italy with funding from the Italian Ministry for University and Research.

An example image is shown in Figure 1. This is the original image as provided to the processing center by a remote operator. Identifying informations have been removed from the image to protect sensitive data, but the image is otherwise as received from the source.

Figure 2 shows the result of applying the roiselect tool. This provides an estimate of the region of interest in the image. The tool implements some heuristics, derived from a-priori informations, about which parts of the image are probably or certainly useless (background, peripheral areas of the image, parts with text) and about how an interesting part should look like. This results from processing several images and obtaining a prototype of the interesting area. For the present study these a-priori informations have been obtained from manually annotated samples of images.

Figure 3 is the output of the roiextract tool, which does nothing but clipping the part of image labeled as the region of interest in the previous step.

The imtclus tool uses the "graded possibilistic fuzzy clustering" [1] method to discover potentially interesting regions in an ultrasound image, and outputs them as prototypes; the prototypes are then labeled as corresponding or not to areas which are actually of interest. A target image is then compared to the prototypes and its pixels are labelled according to the extent to which they are affine to the desired prototypes (fuzzy membership). This information is finally encoded in the image in the form of a color shade, whose intensity corresponds to the level of match between the target pixels and the prototypes.

Figure 4 is the output of the imtclus tool applied to the clipped region of interest. We can see the candidate areas for IMT measurements, which can then be sorted by simple geometrical criteria to eliminate areas which are uninteresting.


[1] F. Masulli, S. Rovetta. Soft Transition from Probabilistic to Possibilistic Fuzzy Clustering. IEEE Transactions on Fuzzy Systems, Vol. 14, No. 4, August 2006, pp. 516-527


DISI - University of Genova
Project PRIN, call 2004, no. 2004062740
Funded by the Italian Ministry of Education, University and Research