Model-based Region-of-interest Selection in Dynamic Breast MRI

Florence Forbes, Nathalie Peyrard, Chris Fraley, Dianne Georgian-Smith, David M. Goldhaber, Adrian E. Raftery
2006 Journal of computer assisted tomography  
Magnetic resonance imaging (MRI) is emerging as a powerful tool for the diagnosis of breast abnormalities. Dynamic analysis of the temporal pattern of contrast uptake has been applied in differential diagnosis of benign and malignant lesions to improve specificity. Selecting a region of interest (ROI) is an almost universal step in the process of examining the contrast uptake characteristics of a breast lesion. We propose an ROI selection method that combines model-based clustering of the
more » ... with Bayesian morphology, a new statistical image segmentation method. We then investigate tools for subsequent analysis of signal intensity time course data in the selected region. Results on a database of 19 patients indicate that the method provides informative segmentations and good detection rates. Key Words: image segmentation, region of interest selection, magnetic resonance imaging, MR mammography, dynamic contrast-enhanced breast MRI, time-signal intensity curves, model-based clustering, Bayesian morphology (J Comput Assist Tomogr 2006;30:675Y687) M agnetic resonance imaging (MRI) is emerging as a powerful tool for the diagnosis of breast abnormalities. Its unique ability to provide morphological and functional information can be used to assist in the differential diagnosis of lesions that other methods find questionable. Many studies have demonstrated the usefulness of MRI in the evaluation of the extent of breast cancer and in treatment planning. It is currently viewed as a complementary diagnostic modality in breast imaging. A number of recent surveys treat breast MRI issues. 1Y5 Because of the high reactivity of breast carcinomas after gadolinium injection, MRI has the potential to allow differentiation between malignant and benign tissues. However, there are as yet no firm standards for data acquisition, post-processing, image analysis, and interpretation of dynamic breast MRI results. It is well known that some benign lesions also enhance, as a result reducing the specificity of MRI. Several methods have been investigated to improve the discrimination between benign and malignant lesions. Lexicons have been designed to standardize the rating and reporting of lesions depicted on magnetic resonance (MR) images and to reduce inter-and intraobserver variability. 6 Improvements have also been achieved through development of contrast agents and pharmacokinetic models. The ultimate goal is to produce sophisticated computer-aided diagnostic tools combining an expanding knowledge base of expert information with stateof-the-art algorithmic techniques for lesion localization, visualization, and classification. 7Y12 In the current study, we focus more specifically on region-of-interest (ROI) selection via dynamic analysis of the temporal pattern of contrast uptake to improve specificity. The criteria that are in use for differential diagnosis can be divided into those related to lesion enhancement kinetics and those related to lesion morphology. Signal intensity time course data are useful for differentiating benign from malignant enhancing lesions. The overall shape of the timesignal intensity curve is an important criterion, whereas a single attribute of the curve, such as the enhancement rate, may not be enough. The evaluation of morphological features and the extraction of architectural information is usually also based on postcontrast images of enhancing areas, integrating qualitative with quantitative diagnostic criteria. Selecting an ROI is an important first step in the process of examining the contrast uptake characteristics of a breast lesion. However, no standard method for ROI selection and analysis of dynamic breast MR data has yet been established. As regards tissue classification, there has been considerable research in brain MRI. Many methods are based on modeling the image intensity with a Gaussian mixture model via the ExpectationYMaximization algorithm. 13 Extensions and variations allow the integration of spatial information into the classification process, using Markov random fields. 14,15 However, there are differences in the analysis of breast MRI and brain MRI, and less research has been devoted to the former. In breast MRI analysis, image segmentation (e.g., finding the ROI) is central. Also breast tissues are much more heterogeneous than brain tissues: normal breasts can consist almost entirely of fatty tissues or include extremely dense fibroglandular tissues, resulting in additional challenges for the analysis of breast MRI. In dynamic breast MRI, the information to be modeled at each ORIGINAL ARTICLE
doi:10.1097/00004728-200607000-00020 pmid:16845302 fatcat:4u2slirtofdwdg5kfjacf6zbbe