-----------------------------------------------------------------------
                    BIOINFORMATICS COLLOQUIUM
               School of Computational Sciences
                    George Mason University
-----------------------------------------------------------------------
                    
            Breast Density and Breast Cancer Risk

                       John Kaufhold, PhD
              SAIC, Advanced Concepts Business Unit


                     Tuesday, April 4, 2006
                             4:30 pm
           Verizon Auditorium, Prince William Campus
ABSTRACT 
The healthy breast is composed of two types of tissue with distinct x-
ray attenuation properties: fibroglandular equivalent (dense) and fatty 
equivalent. Wolfe recognized that the relative amounts and patterns of 
dense tissue was a predictor for breast cancer and proposed a scoring 
system for breast density that's evolved from a subjective ranking by 
radiologists toward a more objective measure of x-ray attenuation. Even 
though the link between breast density and breast cancer risk was 
posited some 30 years ago, estimates of relative risk based on density 
alone still vary widely, and the relationship of breast density to 
other epidemiological risk factors (like family history, e.g.) as well 
as its quantitative response to therapy is still poorly understood 
quantitatively. Here we chronicle in outline the evolution of breast 
density measurements from subjective scores on film-screen mammograms 
to fully volumetric measurements of the proportion of each of 
fibroglandular and fatty tissue in the breast. We rank the measurement 
errors of different phenomena in the newer volumetric methods, from 
compressed breast height estimation error (the largest error source) to 
individual x-ray counting statistics per pixel. An anecdotal 
measurement of year-over-year temporal stability of measurements on a 
digital mammography device illustrates remarkable consistency compared 
to other potential error sources. Fully volumetric digital mammographic 
density measurement results of a pilot longitudinal study of patients 
followed over 3 years trend the same as categorical radiologist density 
rankings. In sum, we diagram a hypothesized (and substantially proven) 
link between breast density and breast cancer risk (via a feedback 
model), but note that getting this model to the clinic still requires a 
good deal of quantification and epidemiological modeling on large 
populations before it will become useful for helping make 
individualized decisions about any specific woman's breast care 
management.

BIOSKETCH
Dr. Kaufhold's education and experience are broadly geared toward 
information extraction from multidimensional signals (images, e.g.) 
embedded in uncertainty (noise) in the form of both structure and 
sensor imperfections. Beginning in 1993, Dr. Kaufhold investigated 
signal processing algorithms for recognition of telephone speech under 
Mari Ostendorf and helped build the WBUR train and test data corpus for 
speech recognition. In 1995, Dr. Kaufhold was awarded a Whitaker 
Fellowship, which he used to study medical image analysis in Boston. In 
that role, he collaborated with: the Neurology Department at 
Massachusetts General Hospital (MGH) on MR brain segmentation and RF 
coil correction, the Boston Heart Foundation on joint segmentation and 
motion estimation of long/short axis ultrasound blood vessel imagery, 
Boston University's Hearing Research Laboratory on 3D depth recovery 
from optically sectioned light microscopy images of cochlear neurons, 
and MIT's Stochastic Signals Group (LIDS) on estimation theoretic 
approaches to image segmentation, and the Surgical Planning Lab (SPL) 
at Brigham and Women's Hospital (BWH) on high-temperature 
superconducting magnets for MR. In 2000, Dr. Kaufhold joined General 
Electric's (GE) Global Research Center, where he focused on 
mammographic imaging (dynamic range management, breast density 
estimation, scatter correction, image formation as well as dual energy 
mammography) and x-ray imaging for interventional cardiac procedures 
(especially stent and guidewire enhancement). Also at the research 
center, Dr. Kaufhold developed machine learning approaches for 
estimating context (roads, grass, trees, vehicles, buildings, and 
shadows) for aerial video analysis, especially as that context 
estimation enables moving object detection and tracking. In March, 
2005, Dr. Kaufhold joined the Center for Pattern Analysis and Decision 
Support Systems in the Advanced Concepts Business Unit at SAIC. Dr. 
Kaufhold's interests include machine learning, recognition from images, 
mammography, and image grammars.

-----------------------------------------------------------------------  
Refreshments are served at 4:00 pm. Find the schedule and directions at 
http://www.binf.gmu.edu/colloq.html