Zero-Shot Learning for Real-Time Ultrasound Image Enhancement

Ultrasound(US) imaging has been widely used for clinical diagnosis. However, ultrasound images inherently suffer from speckle noise and low contrast. Although deep learning-based approaches have been proven to outperform traditional filtering algorithms in image enhancement tasks, they usually require high-quality images for training, which are unavailable in practice. In this paper, we develop a zero-shot learning framework for real-time ultrasound image enhancement. In our framework, relatively low-quality(LQ) and high-quality(HQ) images are beamformed using raw Synthetic Aperture Ultrasound (SAU) data with different channel numbers. The generated LQ-HQ image pairs are used to train two designed U-Net variants whereby one is for inference on GPU and the other with depth-wise separable convolution for inference on CPU and mobile devices. The trained models are directly used for enhancing noisy HQ US images without requiring speckle-free and high contrast targets for training. Experiment results indicate that our method performs more favorably against state-of-the-art image enhancement approaches in terms of both image quality and inference time.