UT researchers are developing novel computer-aided detection and diagnosis strategies (CAD) for breast cancer.
UT researchers are developing novel computer-aided detection and diagnosis strategies (CAD) for breast cancer. Specifically, the research is focused on the automatic detection and diagnosis of spiculated masses, a type of breast cancer finding, commonly observed on mammography (routinely used screening modality for the detection of breast cancer). Collaborators include UT ECE professor Dr. Alan C. Bovik, director of LIVE, UT BME professor Dr. Mia K. Markey, director of BMIL, and researchers at the Department of Diagnostic Radiology, UT M D Anderson Cancer Center.
Spiculated masses are highly malignant and failing to detect these findings early can prove fatal. Unfortunately, detecting these findings is not easy as these masses are invariably submerged in the dense tissue background. Studies have shown spiculated masses account for a fairly large proportion of missed cancers by both radiologists and CAD algorithms.
UT researchers have developed a novel model-based computer-aided detection framework for the detection of spiculated masses on mammography.1 As a part of this framework, they have invented a new class of filters called Spiculated Lesion Filters (SLFs). SLFs are a bank of multi-scale pattern matching filters, which are a new class of complex quadrature filters. What makes the design of these filters novel is that the filters are parameterized from a statistical model of shape measurements of real spiculated masses obtained from mammograms. Their innovative methods for improving the detection performance of SLFs use two ideas: (1) applying a “magnifying glass” to suspect lesions using powerful local analysis techniques and (2) conducting this local analysis by seeking to better match the natural structure and irregular shapes of spicules. Specifically, in order to reduce the false positive rate of this detection algorithm, they have recently pioneered a model-based active contour algorithm called “Snakules.” 2
At each suspect spiculated lesion location that has been optimally identified by the SLFs, “Snakules” are deployed that consist of converging open-ended active linear functions that fall within the category of snakes. The set of convergent snakules has the ability to deform, grow and adapt to the true spicules in the image by an attractive process of curve evolution and motion that optimizes the local matching energy. The candidate points from which spicules originate are automatically detected, and snakules are deployed at these points and made to evolve and grow till the spicule has been completely traced. This evolution creates non-uniformly distributed active snakules having variable lengths and tortuosities that optimally match and “adhere to” the true spicules, from which very specific discriminatory spicule features can be extracted and used in the classification of suspicious locations of spiculated masses detected on mammograms by the SLFs.3
To summarize, UT researchers have pioneered a model-based framework for the detection of spiculated masses on mammography. Unlike any prior methods, models of spiculated lesions were used to construct mathematically optimal pattern matching filters to detect suspect spiculated masses on mammograms. Another completely novel and key contribution of this work is that the approach is model-based, since the parameters of the detection algorithm are derived from measurements, made by experienced radiologists, of physical properties of spiculated masses as seen on mammograms. The University of Texas at Austin has filed a US patent and is communicating with industry partners to license this technology. Their present focus is to lay the groundwork for model-based CAD of breast cancer on 3D x-ray images of the breast. They hypothesize that a 3D, statistical model of the natural variation of spiculated masses will provide the foundation for robust, model-based CAD of breast cancer on 3D x-ray imaging. Model-based algorithms have high potential for generalization because they avoid ad hoc design choices that may be specific to the development images. This is a key point for some of the newer 3D breast imaging modalities such as breast tomosynthesis since early CAD studies of these investigational modalities are inevitably based on a limited number of cases.
For more information on this reasearch, please visit:
Laboratory of Image and Video Engineering (LIVE)
Biomedical Informatics Laboratory (BMIL)
References:
Sampat, M.P., A.C. Bovik, G.J. Whitman, M.K. Markey, A model-based framework for the detection of spiculated masses on mammography. Medical Physics, 2008. 35(5): p. 2110-2123.
Muralidhar, G.S., A.C. Bovik, et al., Snakules – a model-based active contour algorithm for the annotation of spicules on mammography. IEEE Transactions on Medical Imaging, in press.
Muralidhar, G.S., M.K. Markey, and A.C. Bovik. Snakules for automatic classification of spiculated mass locations on mammography. in IEEE Southwest Symposium on Image Analysis and Interpretation. 2010. Austin, TX, USA: IEEE.
