USDL: Inexpensive medical imaging using Deep learning techniques and ultrasound technology

Abstract

In this study, we present USDL, a novel model that employs deep learning algorithms in order to reconstruct and enhance corrupted ultrasound images. We utilize an unsupervised neural network called an autoencoder which works by compressing its input into a latent-space representation and then reconstructing the output from this representation. We trained our model on a dataset that compromises of 15,700 in vivo images of the neck, wrist, elbow, and knee vasculature and compared the quality of the images generated using the structural similarity index (SSIM) and peak to noise ratio (PSNR). In closely simulated conditions, the architecture exhibited an average reconstruction accuracy of 90% as indicated by our SSIM. Our study demonstrates that USDL outperforms state of the art image enhancement and reconstruction techniques in both image quality and computational complexity, while maintaining the architecture efficiency.