[Paper Review] Ultrasound Shear Wave Elasticity Imaging with Spatio-Temporal Deep Learning

Maximilian Neidhardt, Marcel Bengs+5 more|arXiv (Cornell University)|Apr 11, 2022

Ultrasound Imaging and Elastography1 citations

TL;DR

This paper proposes 3D spatio-temporal convolutional neural networks (CNNs) for real-time, pixel-wise elasticity estimation from real ultrasound shear wave data, bypassing conventional velocity-based methods. The method achieves a mean absolute error of 5.01 ± 4.37 kPa on homogeneous phantoms and reduces MAE by 53.93% in inclusions compared to time-of-flight methods, enabling accurate local elasticity estimation independent of push location and even within the push region.

ABSTRACT

Ultrasound shear wave elasticity imaging is a valuable tool for quantifying the elastic properties of tissue. Typically, the shear wave velocity is derived and mapped to an elasticity value, which neglects information such as the shape of the propagating shear wave or push sequence characteristics. We present 3D spatio-temporal CNNs for fast local elasticity estimation from ultrasound data. This approach is based on retrieving elastic properties from shear wave propagation within small local regions. A large training data set is acquired with a robot from homogeneous gelatin phantoms ranging from 17.42 kPa to 126.05 kPa with various push locations. The results show that our approach can estimate elastic properties on a pixelwise basis with a mean absolute error of 5.01(437) kPa. Furthermore, we estimate local elasticity independent of the push location and can even perform accurate estimates inside the push region. For phantoms with embedded inclusions, we report a 53.93% lower MAE (7.50 kPa) and on the background of 85.24% (1.64 kPa) compared to a conventional shear wave method. Overall, our method offers fast local estimations of elastic properties with small spatio-temporal window sizes.

Motivation & Objective

• To develop a deep learning-based method for direct, local elasticity estimation from real ultrasound shear wave data, avoiding reliance on shear wave velocity estimation.
• To improve accuracy and robustness in elasticity imaging, especially in inhomogeneous tissues and near the acoustic push region.
• To enable real-time, pixel-wise elasticity mapping using small spatio-temporal windows without dependence on push location or wave direction.
• To validate the method on both homogeneous phantoms and complex inclusions, demonstrating performance gains over conventional time-of-flight methods.

Proposed method

• The method employs 3D spatio-temporal CNNs that process localized 3D data cubes (spatial x, y, and temporal t) as input, referred to as spatio-temporal windows.
• The network architecture is based on DenseNet blocks with transition layers using average pooling to downsample spatial and temporal dimensions.
• Input data are high-speed ultrasound shear wave sequences from a robotic system, covering a range of elasticities (17.42–126.05 kPa) and push locations.
• The network predicts Young’s modulus directly from the spatio-temporal window, enabling local elasticity estimation at each pixel in the image.
• Training is performed on a large synthetic dataset of homogeneous gelatin phantoms with known elasticities, enabling generalization to inhomogeneous structures.
• Inference is optimized for speed, achieving 0.07 ms per prediction on small windows (0.32 × 0.4 mm), outperforming conventional methods in efficiency.

Experimental results

Research questions

• RQ1 Can 3D spatio-temporal CNNs achieve accurate, local elasticity estimation from real ultrasound shear wave data without relying on shear wave velocity estimation?
• RQ2 How does the proposed method perform in estimating elasticity within the acoustic push region, where conventional methods often fail?
• RQ3 To what extent does the method improve accuracy in inhomogeneous phantoms with stiff inclusions compared to conventional time-of-flight methods?
• RQ4 How does the choice of spatio-temporal window size affect boundary definition and overall accuracy in elasticity estimation?

Key findings

• The proposed method achieves a mean absolute error (MAE) of 5.01 ± 4.37 kPa on homogeneous gelatin phantoms across a wide elasticity range (17.42–126.05 kPa).
• For phantoms with embedded stiff inclusions, the MAE is reduced by 53.93% (to 7.50 kPa) compared to conventional time-of-flight methods.
• On the background tissue, the MAE is 1.64 kPa, representing a 85.24% improvement over time-of-flight methods.
• The method provides consistent elasticity estimates independent of push location and performs accurately even inside the push region.
• Smaller spatio-temporal windows (e.g., 17×17 pixels, ~1×1 mm) yield comparable Dice coefficients to larger windows, but performance degrades significantly below 17×17 pixels.
• The method enables real-time elasticity mapping with inference time as low as 0.07 ms per estimate, outperforming conventional methods in speed and accuracy.