To more fully exploit the basic science of mechanobiology as it pertains to breast cancer progression, the medical imaging field continues to search for fast, safe, and effective elasticity imaging methods. In this project we propose a fundamentally new approach to ultrasonic elasticity imaging in which the weak forces applied to patient tissues and the measured displacements that result are used to train a numerical model specifically for that patient. This constitutive model is developed using finite-element methods and neural networks assembled in a unique configuration called the AutoProgressive (AutoP) Method. AutoP ?learns? complete stress and strain properties directly from sparse force and displacement measurements and without a mathematical model. Using quasi-static stimuli, AutoP exploits the fact that each force-displacement estimate contains information about mechanical properties at all locations in the contiguous tissue. From measurement information and conservation laws, AutoP generates an informational model without the need to make assumptions about tissue linearity, isotropy, or other material properties normally required when constructing images that display tissue mechanical properties. Once an accurate material-property model is formed by AutoP, we adopt a separate rheological model (e.g., Kelvin Voigt) to form viscoelastic parameters for image display. The AutoP method employs beamformed RF-echo acquisitions from which point displacements are estimated, applied compressional force sensed at the transducer surface, and tissue shape. No other imaging method is capable of estimating all relevant stress fields, which gives AutoP unique capabilities. AutoP estimates the mechanical properties that one strives to obtain from an inverse problem approach, but AutoP is not a mathematical inverse technique and hence does not suffer from nonunique solutions. Without the need to assume material properties, AutoP can (in principle) model tissue properties in three dimensions and in time following large deformations in highly-nonlinear, anisotropic media. This R21 proposal focuses on establishing the feasibility of the AutoP methodology for subsequent clinical trials under future funding. At the completion of this two-year project, we will demonstrate a new tool for medical imaging capable of exploring the mechanical properties of tissues over a very broad range of deformations. We specifically target ultrasonic methods and quasi-static force stimuli in this project. However AutoP could eventually be coupled to other imaging modalities (e.g., MRE, OCE) or dynamic force-stimulus methods. The scientific premise underlying AutoP is that we already record all of the information needed to generate a very broad range of elasticity images. The key to unlocking this information is to set aside mathematical models in favor of data-driven informational models built using the unique machine learning abilities of the AutoP method. If successful, AutoP will have a major influence on medical elasticity imaging methods.
A new type of machine-learning method is proposed for elasticity imaging of patient tissues that offers the potential to minimize many current limitations. If we are successful in this development, ultrasonic elasticity imaging would significantly improve in its ability to accurately represent mechanical properties of breast tissues, without any changes in current instrumentation and without increasing cost to the healthcare system or adding risk to patients.
