Ultrasound strain elastography-based characterization of in vivo breast lesions using deep learning

Abstract

Ultrasound strain elastography showed great promise for breast lesion characterization. With the readily available computational power in the clinical workflow, machine learning-based is gaining acceptance in cancer imaging applications. In this study, we combine the latest developments in motion tracking that enables us to obtain high-quality axial strain and full shear strain images to test the deep-learning-based prediction of breast lesion malignancy. To our knowledge, total shear strain images have not been widely used for breast lesion differentiation. More specifically, a PDE-based regularization method [Guo et al., Ultrasonic Imaging (2015)] developed by our group has been improved to get high-quality two-dimensional displacement data. From those high-quality displacement data, total shear strain images can be constructed. We adopted three variants of convolution neural network (CNN) models with attention modules to predict breast lesion malignancy by combining B-mode, axial strain, and total shear strain images and their radiomic features. From our internal database, 150 cases of pathologically-confirmed breast ultrasound data [Hall et al., UMB (2003)] with data augmentation are used to evaluate our machine learning models. Our initial testing results are encouraging with the accuracy and area under the curve being around 0.75.