Echocardiogram enhancement using supervised manifold denoising.

Journal: Medical image analysis
Published Date:

Abstract

This paper presents data-driven methods for echocardiogram enhancement. Existing denoising algorithms typically rely on a single noise model, and do not generalize to the composite noise sources typically found in real-world echocardiograms. Our methods leverage the low-dimensional intrinsic structure of echocardiogram videos. We assume that echocardiogram images are noisy samples from an underlying manifold parametrized by cardiac motion and denoise images via back-projection onto a learned (non-linear) manifold. Our methods incorporate synchronized side information (e.g., electrocardiography), which is often collected alongside the visual data. We evaluate the proposed methods on a synthetic data set and real-world echocardiograms. Quantitative results show improved performance of our methods over recent image despeckling methods and video denoising methods, and a visual analysis of real-world data shows noticeable image enhancement, even in the challenging case of noise due to dropout artifacts.

Authors

  • Hui Wu
    China Medical University College of Health Management, Shenyang 110122, Liaoning Province, China.
  • Toan T Huynh
    Department of General Surgery, Carolinas Medical Center.
  • Richard Souvenir
    Department of Computer Science, University of North Carolina at Charlotte.