Deep-DoPIo: Deep Learning for Shear Elasticity Prediction in Double-Profile Intersection Ultrasound.
Authors
Abstract
We herein propose Deep-DoPIo, a deep learning-based framework that predicts shear elastic modulus from paired narrow- and wide-track displacement profiles generated by a single acoustic radiation force (ARF) push. Unlike the traditional double-profile intersection (DoPIo) ultrasound technique, which relates the intersection time (tint) between these profiles to shear modulus through a linear empirical model, Deep-DoPIo leverages neural networks to learn this mapping directly from the displacement data. Four architectures were evaluated: a feedforward neural network (FFNN), a 1-D convolutional neural network (CNN), an encoder-transformer (E-Tr), and a 1-D vision transformer (ViT). Models were trained on simulated displacement data and fine-tuned on calibrated phantom measurements for domain adaptation. Across simulated materials, traditional DoPIo yielded increasing mean absolute errors (MAEs) with increasing material stiffness, with MAE exceeding 13 for 28-kPa material. Conversely, all examined deep neural network (DNN) models maintained MAEs below 0.5 kPa regardless of material stiffness. In calibrated phantoms, FFNN and CNN produced similar error trends to traditional DoPIo (MAE ≈ 4-8 kPa in a 22.20-kPa medium), indicating limited benefit over the empirical approach. By contrast, transformer-based models (E-Tr and ViT) consistently yielded low errors (<0.2 kPa) across all media of varying stiffness. In a custom liver phantom with a gelatin-graphite inclusion, traditional DoPIo substantially underestimated elasticity relative to shear wave elasticity imaging (SWEI), with errors approaching 50%. In contrast, transformer-based elasticity predictions closely matched those of SWEI, differing by less than 4% in the inclusion and less than 9% in the background. These results demonstrate that transformer-based architectures can generalize from simulation to both calibrated and excised tissue phantoms, providing accurate shear modulus estimates under conditions where traditional DoPIo struggles. Future work will evaluate performance in vivo and systematically compare Deep-DoPIo against SWEI under clinically relevant conditions.