A novel CycleGAN network applicable for enhancing low-quality ultrasound images of multiple organs

Medical ultrasound imaging is one of the most commonly used imaging modalities for clinicians to make clinical diagnoses. Currently, medical ultrasound imaging equipment is transitioning towards portable ultrasound devices that can be more widely used. However, portable ultrasound devices suffer from the drawback of low imaging quality due to hardware limitations. Enhancing ultrasound images requires paired medical ultrasound image data with identical content but different styles for enhancement training, which is quite challenging. To address these issues, this paper proposes a DPM CycleGAN, a generative adversarial network that can convert low-quality ultrasound images into high-quality ones without the need for paired data. This network employs a maximum perception feature extraction module designed specifically for ultrasound images to extract global features. Subsequently, a diffusion feature enhancement module integrated with pseudo-labels is used to enhance these global features. The generator of DPM CycleGAN utilizes the enhanced global features to generate high-quality medical ultrasound images. Experimental results on the USenhance 2023 ultrasound dataset demonstrate that this method can effectively enhance low-quality ultrasound images of various organs, not only improving the resolution of ultrasound images but also enhancing the organ structure and texture information, and eliminating some noise in the ultrasound images. Compared with the latest ultrasound image enhancement methods, the model proposed in this paper performs optimally in all metrics, which fully demonstrates the promising potential of the proposed method in medical image enhancement and its valuable support in areas such as disease prediction and treatment.