WSRGAN: A wavelet-based GAN for super-resolution of plane-wave ultrasound images without sampling loss

Abstract

Plane-wave ultrasound imaging has ultra-fast temporal resolution and can image at frame rates exceeding 1 kHz, enabling breakthrough technologies like shear wave elastography. However, the imaging process of this technology lacks focus and poor imaging quality and usually requires beamforming technology to assist imaging, such as Coherent Plane-Wave Compounding (CPWC). The imaging quality of this method is often directly related to the number of beams, sacrificing the time advantage of single-beam plane wave imaging. This study proposes a Wavelet-based Generative Adversarial Network for Super-Resolution (WSRGAN) to meet the growing demand for high-quality imaging of single-beam plane waves. WSRGAN uses the encoding and decoding network as the generator and uses wavelet transform to replace the sampling process. Furthermore, we further designed adaptive wavelet attention to enable the model to focus attention on different levels of features at different stages. Wavelet GAN loss was proposed as GAN loss, and a new combined loss function was designed for the generator. The experiment was conducted on the Plane-wave Imaging Challenge in Medical UltraSound (PICMUS) 2016 dataset. On the point target, the Full Width at Half Maximum (FWHM) of WSRGAN reached 0.268 mm to 0.502 mm. On the cyst target, the contrast of WSRGAN reaches 26.156 dB to 37.223 dB. Among invivo targets, the Peak Signal-to-Noise Ratio (PSNR) of WSRGAN is 41.518 dB, and the Structural Similarity Index Measurement (SSIM) is 0.984. Experiments show that each module proposed by the model has a total contribution and is helpful for various experiments. All targets performed well.