Emulating Clinical Quality Muscle B-mode Ultrasound Images from Plane Wave Images Using a Two-Stage Machine Learning Model

TL;DR

This work addresses the gap between research ultrasound images and clinical-quality B-mode images for muscle, particularly under plane-wave imaging. It introduces a two-stage ML pipeline: Stage 1 a U-Net learning plane-wave compounding and a traditional image-processing pipeline, and Stage 2 a CycleGAN that translates to clinical-style muscle images, all deployed in real time on a Verasonics Vantage system. Stage 1 improves similarity to ground truth with lower RMSE and higher SSIM, while Stage 2 reduces speckle and increases CNR and fiber cohesion, validated by a reader study. The combined approach achieves real-time enhancement at about 28.5 FPS, offering a practical path to real-time, clinically styled muscle ultrasound during SWEI acquisitions and potential improvements in downstream analyses like muscle segmentation and fiber orientation.

Abstract

Research ultrasound scanners such as the Verasonics Vantage often lack the advanced image processing algorithms used by clinical systems. Image quality is even lower in plane wave imaging - often used for shear wave elasticity imaging (SWEI) - which sacrifices spatial resolution for temporal resolution. As a result, delay-and-summed images acquired from SWEI have limited interpretability. In this project, a two-stage machine learning model was trained to enhance single plane wave images of muscle acquired with a Verasonics Vantage system. The first stage of the model consists of a U-Net trained to emulate plane wave compounding, histogram matching, and unsharp masking using paired images. The second stage consists of a CycleGAN trained to emulate clinical muscle B-modes using unpaired images. This two-stage model was implemented on the Verasonics Vantage research ultrasound scanner, and its ability to provide high-speed image formation at a frame rate of 28.5 +/- 0.6 FPS from a single plane wave transmit was demonstrated. A reader study with two physicians demonstrated that these processed images had significantly greater structural fidelity and less speckle than the original plane wave images.

Figures (10)

• Figure 1: Left column: Single plane wave images taken along the transverse and longitudinal views of the vastus lateralis. Right column: Single plane wave images after processing by MimickNet.
• Figure 2: Example images from clinical datasets used to train the CycleGAN. Image A is of the gastrocnemius from dataset 1, image B is of the medial gastrocnemius and soleus from dataset 2, images C and D are of the biceps brachii and medial gastrocnemius respectively from dataset 3, and image E is of the vastus lateralis from dataset 4. All datasets are described in Table \ref{['tab:1']}.
• Figure 3: CycleGAN architecture. The first stage U-Net shares the same architecture as the CycleGAN’s generator.
• Figure 4: Setup used to image a healthy volunteer using the Verasonics Vantage™ system and L7-4 transducer.
• Figure 5: Example of ROIs used to calculate image metrics. Standard deviation was calculated over the red box to quantify speckle. CNR was calculated over the red and blue boxes. Standard deviation was calculated over the red line segment to characterize fiber and fascicle cohesiveness.