Program Information
Improved Medical Image Denoising by Stochastic Global Patch Search
D Karimi1*, D Ruan2 , A Sawant3 , (1) ,,,(2) UCLA School of Medicine, Los Angeles, CA, (3) University of Maryland School of Medicine, Baltimore, MD
Presentations
SU-J-CAMPUS-IT-1 (Sunday, July 30, 2017) 4:00 PM - 5:00 PM Room: Imaging ePoster Theater
Purpose: Effective denoising of medical images can greatly improve their diagnostic value. Many of the published denoising algorithms use nonlocal patches of high intensity similarity for enhancement. However, typically a small local window is searched for similar patches. The goal of this study was to investigate the impact of adopting a global search strategy on the denoising performance.
Methods: We develop a stochastic global search algorithm for finding similar patches in 3D medical images. Our algorithm exploits inter-slice similarities to efficiently find patch match patterns. We apply a collaborative filter to the similar patches for denoising. In order to assess the advantage gained by global patch search, we compare computed tomography (CT) images denoised using local and global search strategies in terms of the root-mean-square of the error (RMSE) and the visual quality of image features.
Results: For target patches that contain features such as edges or texture, a local search fails to find a large number of similar patches, while our proposed algorithm can efficiently find similar patches. This significantly improves the visual quality of the fine features in the denoised image. It also decreases the RMSE compared with images denoised using a local search (26.7 versus 26.9), even though the difference is not statistically significant (p= 0.10).
Conclusion: A global search is necessary for finding similar patches for patterned target patches that usually contain the most important image features. The similarity of neighboring slices in 3D medical images can be used to guide the patch search. Collaborative filtering of these patches leads to superior denoising results, effectively suppressing the noise while preserving the genuine image features.
Funding Support, Disclosures, and Conflict of Interest: This work is supported in part by NIH 5R01CA169102-05
Contact Email: