Magnetic resonance imaging (MRI) is a revolutionary imaging technology for many medical applications. The diagnosis of many forms of cancer and other serious conditions is made possible without invasive surgery. However, there is a significant computational burden involved in the production and presentation of MRI data for diagnostic purposes. The fundamentally different principles in volume-based imaging, and in MRI in particular, have motivated the development of a wide range of new algorithms and techniques.

In this short paper, Song et al. present a pair of techniques: one for the segmentation of breast MRI data and one for the visualization of such segmented breast image data. The segmentation task is to accurately distinguish fat cells from fibroglandular tissue and air. The authors briefly mention a variety of other segmentation techniques, most notably fuzzy clustering methods and semiautomated approaches requiring manual initialization or guidance. It should be noted that there is work related to segmenting skin regions in breast MRI data, but this is not the emphasis of the current paper.

The authors adopt a kernel-based fuzzy c-means method using a Gaussian kernel for the segmentation task. Though no rationale is given for the selection of a Gaussian kernel, it results in a reasonably simple objective function to be minimized. One attribute of the fuzzy c-means method is that its performance can degrade significantly if an inappropriate assumption is made for the number of cluster centers. Breast MRI data with a small, known number of tissue types is well suited for this method. After segmentation is completed, the resulting image data can be visualized by a variety of methods. Song et al. use a straightforward approach that differs little from the general concepts presented in the literature (see, for example, Kuchment [1]). The resulting volumetric visualizations give a readily apparent description of the breast.

For techniques of this sort, an evaluation of the results can be difficult. The authors present a small set of results in image form; in these cases, the segmentation is certainly appropriate. They compare the results subjectively to other techniques in the text of the paper. One interesting metric is the time required for the segmentation to converge; these times range from six to 32 minutes on a reasonably powerful desktop computer, for realistic choices of data size. Convergence times varied depending on initial cluster center assignments; careful selection outperformed random assignment, as would be expected.

This paper is fairly well written and references a few of the most relevant published works on the topic. The methods presented are interesting extensions of known techniques, and the authors show good results from their application. This paper is not intended to--and does not--present a major finding in the application area of breast MRI data processing, but instead presents an interesting technique for such data.