Purchase this article with an account.
Guillaume Gisbert, Neel Dey, Hiroshi Ishikawa, Joel Schuman, James Fishbaugh, Guido Gerig; Improved Denoising of Optical Coherence Tomography via Repeated Acquisitions and Unsupervised Deep Learning. Invest. Ophthalmol. Vis. Sci. 2020;61(9):PB0035.
Download citation file:
© ARVO (1962-2015); The Authors (2016-present)
Optical Coherence Tomography (OCT) is widely used, yet its interpretation is confounded by strong speckle noise. Previous denoising methods make strong assumptions on noise characteristics and struggle to retain fine structure. We present a data-driven registration and denoising method which vastly improves image fidelity while requiring only repeated noisy acquisitions of individual subjects.
Noise2Noise is an unsupervised denoising approach for repeatedly acquired noisy images (Lehtinen, ICML, 2018). By training a convolutional neural network on noisy image pairs with zero-mean noise, the mean of the clean image distribution is learned. However, repeats are assumed to vary only in noise and not in structure, and is thus inapplicable as repeated OCT scans show strong linear and subtle nonlinear mis-alignment.To this end, we estimate subject-wise image templates which minimize geometric deformation to individual repeats (Avants, NeuroImage, 2004). We then non-linearly transform each repeat to its template, thus achieving spatial alignment and allowing the use of Noise2Noise.24 subjects were imaged on a Cirrus HD5000 with 600 XY slices, with each undergoing 6 repeats, yielding 30 pairs of noisy repeats per subject. After alignment, we train a 2D U-Net slice-by-slice on the dataset of 432,000 noisy pairs. Once trained, we denoise each scan by applying the model slice-wise.
Signal-to-noise ratio is inapplicable here as clean reference scans do not exist. We thus present the contrast-to-noise ratio (CNR) on a pair of regions-of-interest. In Fig.1, the raw scan shows a CNR of 1.76, whereas our denoised image achieves 126.22, a ~71x improvement. By comparison, the popular unsupervised BM4D denoiser achieves a CNR of 16.60.
Through an unsupervised joint registration and denoising solution, we improve OCT images both qualitatively and quantitatively, allowing for clearer morphological interpretation. While we leverage repeated observations for training, the model can be directly applied to unseen individual scans from new subjects thereafter, making it applicable to large amounts of retrospective clinical data.
This is a 2020 Imaging in the Eye Conference abstract.
A collage of results. Rows correspond to different views. Col. 1: A single scan. Col. 2: Results of using the BM4D denoising method. Col. 3: A linear average of 6 repeated scans. Col. 4: Results of our registration and denoising framework.
This PDF is available to Subscribers Only