Program Information
Adaptive Medical Image Denoising Over Multiple Anatomical Regions with Edge and Texture Preservation
D Floros1*, AS Iliopoulos2 , N Pitsianis1,2 , X Sun2 , L Ren3 , (1) Aristotle University of Thessaloniki, Greece, (2) Duke University, Durham, NC, (3) Duke University Medical Center, Durham, NC
Presentations
WE-G-201-6 (Wednesday, August 2, 2017) 4:30 PM - 6:00 PM Room: 201
Purpose: To develop a novel robust method for medical image denoising that uses prior anatomical structure information to perform spatially variant noise reduction in each anatomical region in the image, without degrading edge and texture information between and within anatomical structures.
Methods: We introduce a new denoising method that is edge- and texture-preserving, and adaptive to spatial variation in noise behavior. It also has the potential to be computationally efficient. The method, which we refer to as ETA-means, is influenced particularly by Bilateral Filtering and Non-Local Means, and makes novel improvements upon state-of-the-art filtering techniques. A plain image patch centered at each pixel is used for texture description and discrimination. An additional weight factor is used for edge detection and preservation. ETA-means makes use of available prior anatomical structure information to separate regions with different noise and signal characteristics, for two benefits: reduction of computation cost and elimination of spurious feature matching between different anatomical structures. Parallel CPU/GPU architectures are utilized to accelerate the denoising process. The method was evaluated with clinically acquired cone-beam projections of lung cancer patients. Anatomical structures contoured by physicians in the planning CT images were used as prior anatomical information.
Results: ETA-means is robust to segmentation errors. Results are compared with the commonly used implementation of Non-Local Means in the OpenCV library. ETA-means preserves structure while better discriminating texture, and retains clear edges, without blur or reduced contrast. Residual images show obvious structural patterns for the NLM method, whereas minimal structural patterns are observed for the ETA-means method.
Conclusion: ETA-means has evident advantages in denoising medical images, as shown in the clinical data. The method can have wide applications for 2D image such as mammography and chest X-rays, and 3D imaging such as CBCT for noise reduction with edge and texture preservation.
Funding Support, Disclosures, and Conflict of Interest: This research was supported in part by NIH grant No. R01-CA184173, ARO grant W911NF-13-l-0344, and an NVIDIA academic research equipment grant.
Contact Email: