Feature Of The Week 6/24/12: Duke University Researchers Develop Techniques for Denoising OCT Images And Publically Share Their Software
Duke University researchers have a long history of major advances to the field of Optical Coherence Tomography dating back to some of the first OCT work in the very early 1990s by Dr. Joseph Izatt. Since the OCT News website started in 2007 Duke has published over 217 papers that appear on OCT News (See Organizations in the News). One recent example of their work was on “Sparsity Based Denoising of Spectral Domain Optical Coherence Tomography Images”. Below is a summary of this interesting work and along with a link to software they have developed and made publically available.
In this work, we make contact with the field of compressive sensing and present a development and generalization of tools and results for reconstructing irregularly sampled tomographic data. In particular, we focus on the spectral domain optical coherence tomography (SDOCT) systems, and show the applicability of our algorithm in reducing speckle noise.
Previous works in removing speckle noise can be categorized into two groups: model-based single-frame or multi-frame averaging techniques. The first group utilizes classic denoising methods, which often assume an a priori parametric or non-parametric model for the signal and noise (e.g. Wiener filtering or wavelets). The second group relies on capturing a sequence of repeated B-scans from a unique position. In a post-processing step, these images are registered and averaged, to create a less noisy image. The shortcoming of the first group of denoising methods is the introduction of excessive blurring artifacts in the denoised images, while the second group (averaging based denoising) dramatically increases the image acquisition time.
Alternatively, we present a novel hybrid approach, which we call multiscale sparsity based tomographic denoising (MSBTD) for denoising volumetric SDOCT scans. The MSBTD method utilizes a non-uniform scanning pattern, in which, a fraction of B-scans are captured slowly at a relatively higher than nominal signal-to-noise ratio (SNR). The rest of the B-scans are captured fast at the nominal SNR. Utilizing compressive sensing principles, we learn a sparse representation dictionary for each of these high-SNR images and utilize these dictionaries to denoise the neighboring low- SNR B-scans. The rationale for this approach is that neighboring B-scans, in common SDOCT volumes, are expected to have similar texture and noise pattern. The exciting property of our approach is that it does not require capturing more than one B-scan from the majority of azimuthal positions and therefore, it requires significantly less scanning time. Moreover, the MSBTD method is well suited to retrospectively improve the quality of images in databanks from large scale multi-center clinical trials. Note that, traditionally, ocular SDOCT imaging protocols often require at least two types of SDOCT scans: the first, a densely sampled (but low-SNR) large field-of-view volumetric scan and the second, sparsely sampled high-SNR scans (often consisting of only one image from the fovea).
We tested our algorithm on randomly selected data sets from 17 eyes from 17 subjects with and without nonneovascular age-related macular degeneration (AMD) enrolled in the multicenter A2A SDOCT study. All data sets were captured before the start of this project and the imaging protocol was not altered in any form for this study. We demonstrated that the MSBTD method outperforms state-of-the-art denoising methods including: Tikhonov, New SURE, K-SVD, and BM3D methods. To encourage further research in this area, we have made the data set and the software that we have developed for this project along with a high resolution PowerPoint presentation publically available by clicking HERE.
For more information see recent Article. Courtesy of Sina Farsiu from the Duke University.