Signal Processing Seminar

Title:  Ultrasonic Model-based Imaging and Breast Cancer Detection

Speaker: Professor Michael Oelze
Electrical and Computer Engineering
University of Illinois

Date:  Wednesday, February 27, 2008
Time:  4:00 - 5:00 pm
Where:  2269 Beckman Institute

Abstract:  Early detection and diagnosis of breast cancer through imaging leads 
to improved prognosis. Quantitative ultrasound (QUS) imaging techniques were 
explored for classifying tumors in rodent models of breast cancer. The QUS 
imaging technique was based on estimates of scatterer properties obtained through 
parameterizing the ultrasonic backscatter from tissues based on generic models for 
scattering. Initial QUS analysis yielded moderate success for classifying tumors. 
However, in most cases the scatterer property estimates did not correspond to the 
underlying structure it was intended to model. A new technique was developed, 
which allowed models of scattering to be constructed directly from underlying 
tissue microstructure. The new modeling technique was based on the construction 
of three-dimensional impedance maps (3DZMs) of histological sections of tissues. 
Tissue blocks were fixed, stained, sectioned into serial slides, and photographed at 
high magnification. The serial slides were registered to form a three-dimensional 
block. The 3DZM was constructed by associating pixel intensity values to the 
characteristic acoustic impedance of tissues. From the 3DZMs scatterer properties 
could be estimated using existing models, new models could be deduced, and 
scattering sources identified. Scatterer properties estimated from the 3DZMs were 
compared to scatterer property estimates from ultrasonic backscatter. Estimates of the 
average scatterer diameter using the 3DZM and ultrasonic backscatter were within 10% 
relative error. 3DZMs were also used to predict scattering sources and create new 
models for scattering from cells. New models for scattering, the addition of new 
parameters, and more appropriate analysis bandwidths yielded significantly improved 
classification of different kinds of tumors.  Feature analysis plots indicated cancer 
classification was improved when using multiple parameters over few parameters.