Abstract: Aiming at the problems that portable ultrasound devices suffered from low imaging quality due to hardware constraints, high- and low-quality ultrasound images were difficult to pair, and traditional enhancement methods failed to sufficiently preserve fine details, we proposed a registration-free generative adversarial network for the stable enhancement of unpaired medical ultrasound images. The network was based on the cycle-consistent generative adversarial network, we  designed a maximum  perception  feature extraction  enhancement module to extract  global and local features of images while preserving their spatial structures. 
We constructed  a clustering feature enhancement learning network to strengthen the ability of extracting complex details from the image. Meanwhile, the loss function was optimized by integrating pseudo-label classification loss and multilayer perceptron feature loss. Experimental results on the USenhance 2023 ultrasound image enhancement challenge dataset show that the three core evaluation metrics of the proposed network are significantly superior to those of mainstream and state-of-the-art ultrasound image enhancement methods. The enhanced images exhibit lower noise, with clearer structural, textural, and contour details of organs.

Key words: medical ultrasound image enhancement, deep learning, generative adversarial network