Click here to

Session: Quantitative Ultrasound and Emergent Imaging Technology (II) [Return to Session]

High-Resolution Ultrasound Imaging Through Self-Supervised Learning

X Dai*, Y Lei, T Wang, M Axente, J Roper, D Xu, J Lin, J Bradley, T Liu, X Yang, Emory Univ, Atlanta, GA


WE-F-TRACK 3-1 (Wednesday, 7/28/2021) 4:30 PM - 5:30 PM [Eastern Time (GMT-4)]

Purpose: This study aims to develop a deep learning-based algorithm to reconstruct high-resolution 3D ultrasound (US) images from sparsely acquired 2D images without any extra atlas database.

Methods: In conventional US imaging, in-plane spatial resolution is generally much higher than through-plane resolution. Under the assumption of isotropy of US imaging in biological tissue, the mapping from through-plane low-resolution to high-resolution images should be identical as that for in-plane images. Therefore, the mapping can be learned using paired low-resolution and high-resolution in-plane images. Given the acquired 2D images which have high in-plane spatial resolution, the in-plane low-resolution images can be obtained through down-sampling. In this study, a model which is constructed by two independent cycle consistent generative adversarial networks (cycleGANs) is built to learn the mapping from low-resolution to high-resolution images and generate high-resolution through-plane images from original sparely distributed 2D images. We performed a leave-one-out cross validation method to evaluation the proposed high-resolution method using a 3D breast database with 70 patients. Metrics including mean absolute error (MAE), peak signal-to-noise ratio (PSNR), and visual information fidelity (VIF) were computed to quantify the performance of our proposed method.

Results: In the experimental studies with a spatial resolution enhancement factor of 3, for all patients, a mean and standard deviation of MAE value of 0.90±0.15, a PSNR value of 37.88±0.88 dB, and a VIF value of 0.69±0.01 were achieved by our proposed method, significantly outperforming the commonly used bicubic interpolation.

Conclusion: In this study, a novel self-supervised learning framework has been investigated for reconstructing high-resolution 3D US images from sparely acquired 2D images. Without using any extra atlas images, our proposed method achieves significant improvement on through-plane resolution. Its self-supervision capability could accelerate high-resolution US imaging, and this imaging tool will have a high impact on US-guided intervention.



    Not Applicable / None Entered.


    Not Applicable / None Entered.

    Contact Email