Controllable Quality Improvement of Mobile Ultrasound

Mobile ultrasound imaging, widely used in clinical diagnosis due to its non-invasive, convenient, and rapid characteristics, faces challenges such as low image contrast, various artifacts, and noise interference, demanding high clinical experience from physicians. Utilizing image super-resolution technology to enhance ultrasound image quality holds significant clinical importance. Unlike natural scene images, ultrasound images lack paired high and low-resolution datasets, making the reconstruction task more challenging. This paper conducts an in-depth study on the super-resolution reconstruction of ultrasound images using Generative Adversarial Networks (GANs). It performs extensive experiments on a dataset accumulated at the Hospital, achieving favorable results. The main contributions of this paper are as follows: 1) Optimizing the CycleGAN network's generator with hyperparameter convolution kernels to extract multiscale features from input images and train them in combination with hyper-parameters. The super-resolution (SR) images at different scales are reconstructed by a decoder controlled by hyperparameter prompts. 2) Employing both paired simulated and unpaired clinical ultrasound images and natural images to train a super-resolution model that acquires ultrasound-specific features based on natural cognition. 3) Various random damage techniques are applied by employing dynamic random image damage methods such as Gaussian, grid, mosaic, directional, etc. These damage methods can simulate low-resolution when imaging with a mobile ultrasound device based on a limited unpaired dataset, providing many paired samples.