Factors determining the quality of learned elastic-modulus images from ultrasonic measurements

The autoprogressive (AutoP) method forms images of the stress and material properties of tissues using surface force and internal displacement measurements from quasi-static compression experiments. The method combines measurements with finite-element analysis to obtain physically constrained stress-strain training data, which are used to train a neural network material model. Internal displacements must be acquired using a medical imaging modality to apply AutoP to medical tasks. Ultrasound (US) linear arrays are well-suited for two-dimensional displacement tracking, however, tissue deformation occurs in three-dimensional displacement tracking, and displacement quality is limited. This report employs gelatin phantoms to examine the key experimental variables that are responsible for determining image contrast, accuracy, and spatial resolution of elastic-modulus images formed using AutoP with measurements from US linear array transducers. The greatest degradation of image quality occurs with displacement errors and when displacement sampling does not extend over the volume of deformed media. Rather than producing the random noise patterns often observed in images, measurement errors reduce image contrast and spatial resolution. Losses are minimized by techniques that increase the information content of the training data at the cost of extending the training time.