We are developing ultrasonic-based techniques for elastic modulus imaging throughout a tissue volume using the autoprogressive (AutoP) method with linear array transducers. AutoP discovers the constitutive properties of media by estimating the stress and strain in a volume from force-displacement measurements recorded in planes. AutoP is a data-driven machine learning technique that combines shallow neural network structures with object-specific measurements, in stark contrast with traditional deep learning techniques. In this paper, we outline a strategy for acquiring measurements from a series of compression planes across the volume and applying them in a training sequence that efficiently trains networks to model volumetric deformation patterns. We show with phantom studies that measurements collected while compressing the medium in parallel planes throughout the volume can yield elastic modulus image values within 10% of values measured independently. Deformation models are developed accurately in minutes using a comprehensive set of exact measurements. However, experimental limitations slow the learning process and ultimately limit the contrast and spatial resolution of modulus images in ways that can be minimized. An efficient imaging strategy balances the need to provide more planes of force-displacement measurements to enhance learning with the need to manage measurement errors. Test statistics obtained from the developing model can guide the learning process.
Newman et al. (Mon,) studied this question.