A machine learning alternative to model-based elastography

Abstract

Model-based elastography methods suffer severe limitations in imaging the complex mechanical behavior of real biological tissues. We adapted the Autoprogressive Method (AutoP) to address these limitations by approaching the inverse problem with machine learning tools. AutoP combines finite element analysis (FEA) and artificial neural networks (ANNs) with force and displacement measurements to develop soft-computational models of mechanical behavior. Unlike model-based elastography methods, only measurement data inform the material properties learned by the ANNs. Because this machine learning approach foregoes the initial model assumption, AutoP can be applied to anisotropic, time-varying, and nonlinear media common in biomedical imaging applications. We first implemented AutoP to characterize linear-elastic gelatin phantoms and ex vivo rabbit kidneys to demonstrate the potential for medical imaging. Those models required an estimate of the interior geometry via segmentation of the B-mode images. In our current work, the capabilities of AutoP are extended by developing a novel ANN architecture. Incorporating spatial information as part of the input to a pair of ANNs working in tandem allows the models to learn the spatially varying mechanical behavior, thus precluding the segmentation requirement. We will demonstrate this new approach to elasticity imaging by presenting elastograms generated by trained ANN material models.