Deep learning-based image enhancement in optical coherence tomography by exploiting interference fringe

Abstract

Optical coherence tomography (OCT), an interferometric imaging technique, provides non-invasive, high-speed, high-sensitive volumetric biological imaging in vivo. However, systemic features inherent in the basic operating principle of OCT limit its imaging performance such as spatial resolution and signal-to-noise ratio. Here, we propose a deep learning-based OCT image enhancement framework that exploits raw interference fringes to achieve further enhancement from currently obtainable optimized images. The proposed framework for enhancing spatial resolution and reducing speckle noise in OCT images consists of two separate models: an A-scan-based network (NetA) and a B-scan-based network (NetB). NetA utilizes spectrograms obtained via short-time Fourier transform of raw interference fringes to enhance axial resolution of A-scans. NetB was introduced to enhance lateral resolution and reduce speckle noise in B-scan images. The individually trained networks were applied sequentially. We demonstrate the versatility and capability of the proposed framework by visually and quantitatively validating its robust performance. Comparative studies suggest that deep learning utilizing interference fringes can outperform the existing methods. Furthermore, we demonstrate the advantages of the proposed method by comparing our outcomes with multi-B-scan averaged images and contrast-adjusted images. We expect that the proposed framework will be a versatile technology that can improve functionality of OCT.

Generalized 3D registration algorithm for enhancing retinal optical coherence tomography images

“…The MSE has been extensively utilized in OCT studies for this purpose, reflecting its relevance and applicability to the field. 29 , 44 It is the average squared difference between averaged and reference OCT intensity images, defined as where denotes summation over the total number of pixels ( ), and and represent the pixel intensity of the averaged and reference images, respectively. 29 …”

“…The MSE has been extensively utilized in OCT studies for this purpose, reflecting its relevance and applicability to the field. 29,44 It is the average squared difference between averaged and reference OCT intensity images, defined as E Q -T A R G E T ; t e m p : i n t r a l i n k -; e 0 0 3 ; 1 1 7 ; 3 0 8…”

Artificial neural network for enhancing signal-to-noise ratio and contrast in photothermal optical coherence tomography

“…For example, for atomic force microscopy (AFM), Borodinov et al 23 improved the detection limit by more than an order of magnitude using a hybrid deep learning model. In OCT, deep learning (DL) models have been used for various purposes, such as segmentation 24 – 26 and reconstruction of OCT structural tomograms 27 , 28 , dispersion compensation 29 , diagnosis 30 , and classification of retinal disease 31 , 32 , or automated noise and artifact removal 33 , 34 . Deep learning also has been used in functional extensions of OCT.…”

Speckle contrast reduction through the use of a modally-specific photonic lantern for optical coherence tomography

“…Due to its coherent light detection scheme, OCT suffers from speckle patterns that can mask tissue structures [ 2 – 4 ]. Several techniques have been previously proposed to reduce speckle noise [ 5 – 36 ]. These techniques however either compromised image spatial resolution [ 5 – 18 ], required multiple illumination and detection angles for limited improvement [ 19 – 23 ] or can blur intrinsic speckle-size structures of the image by removing speckles [ 24 – 34 ].…”

“…These techniques however either compromised image spatial resolution [ 5 – 18 ], required multiple illumination and detection angles for limited improvement [ 19 – 23 ] or can blur intrinsic speckle-size structures of the image by removing speckles [ 24 – 34 ]. Neural network methods in OCT are emerging and can also be used to reduce speckle noise [ 35 , 36 ]. While these methods are performant, they are limited by the need for available training datasets and may blur intrinsic structures as previously described.…”