Ultrasound Imaging Improved by the Context Encoder Reconstruction Generative Adversarial Network

Abstract

Portable ultrasound devices with less numbers of transducers become popularly spread soon or later. Accordingly, this work explores how to make quality of ultrasonic B-mode images from 32 channels approaching to that from 128 channels by using Deep Neural Networks (DNNs). Based on the concepts of auto-encoder, context encoder, semi-supervised latent-class learning, and context-conditional Generative Adversarial Network (GAN), we develop the Context Encoder Reconstruction GAN (CER-GAN)., which adopts the contour learning mechanism to establish the relationship between 32-channel and 128-channel Point Spread Functions (PSF) via the mini-batch training. In our experiments, 20 pairs of 32 and 128-channel ultrasound images were used where the way of “leave one person out” was employed for cross-validation. Our CER-GAN automatically encodes 32-channel PSF blocks, and generates pseudo 128-channel PSF ones which are matched with real 128-channel ones to obtain discrimination and amendment from the adversarial mechanism. Finally, these pseudo blocks are used to yield an improved image. The experimental results reveal that the proposed CER-GAN outperforms the conventional DNNs. In Full Width at Half Maximum (FWHM), the minimum distinguishable pixel lengths from the 32-channel, CER-GAN, and 128-channel images are 13.34, 11.15, and 8.62, respectively. In Contrast Noise Ratio (CNR)/PICMUS CNR, the scores of the 32-channel, CER-GAN, and 128-channel images are 0.939/2.381, 1.508/6.502, and 1.422/6.002, respectively. The reason of the favorable scores from CER-GAN is that the background speckle noise is lowered to some extent. Therefore, our CER-GAN can effectively ameliorate the quality of 32-channel ultrasound imaging, close to that of 128-channel one.