Blind super-resolution for handheld ultrasound image: Two-stage degradation based unpaired deep learning

Handheld ultrasound devices are widely used in clinical diagnostics and examinations due to their portability. However, their imaging quality is often inferior to that of large-scale ultrasound devices due to hardware limitations.

To enhance the image quality of handheld ultrasound devices, a blind super-resolution method based on two-stage degradation is proposed. The first degradation stage, referred to as frequency probabilistic degradation, is designed to mitigate the structural distortion and texture loss commonly introduced by general probabilistic degradation. In this stage, high-quality ultrasound images acquired from large-scale ultrasound devices are decomposed into high-frequency and low-frequency components using wavelet transform. These two components are respectively processed with blur kernels and noise, both generated by neural networks, and then recombined to produce synthetic images. In the second degradation stage, Gaussian blur kernels and speckle noise are randomly generated and applied to the synthetic images, further degrading their quality and enhancing the diversity of the training samples. Additionally, recognizing that the general perceptual loss function is insufficient to capture the unique characteristics of ultrasound images, a new ultrasound perceptual loss function is introduced.

Eventually, supervised learning is performed using the EDSR model on the synthetic images after two-stage degradation and high-quality images, and blind super-resolution of low-quality ultrasound images is realized.

Experiments are carried out on public datasets to demonstrate the proposed method, the experimental results show that the proposed method outperforms state-of-the-art techniques in terms of image quality improvement.