Background and objectives: Optical coherence tomography (OCT) is a cross-sectional imaging method utilizing a low coherence interferometry. The lateral resolution of the OCT is limited by the numerical aperture (NA) of the imaging lens. Using a high NA lens improves the lateral resolution but reduces the depth of focus (DOF). In this study, we propose a method to improve the lateral resolution of OCT images by end-to-end training of a deep 1-D deconvolution network without use of high-resolution images.
Materials and methods: To improve the lateral resolution of the OCT, we trained the 1-D deconvolution network using lateral profiles of OCT images and the beam spot size. We used our image-guided laparoscopic surgical tool (IGLaST) to acquire OCT images of nonbiological and biological samples ex vivo. The OCT images were then blurred by applying Gaussian functions with various full width half maximums ranging from 40 to 160 µm. The network was trained using the blurred OCT images as input and the non-blurred original OCT images as output. We quantitatively evaluated the developed network in terms of similarity and signal-to-ratio (SNR), using in-vivo images of mesenteric tissue from a porcine model that was not used for training. In addition, we performed knife-edge tests and qualitative evaluation of the network to show the lateral resolution improvement of ex-vivo and in-vivo OCT images.
Results: The proposed method showed an improvement of image quality on both blurred images and non-blurred images. When the proposed deconvolution network was applied, the similarity to the non-blurred image was improved by 1.29 times, and the SNR was improved by 1.76 dB compared to the artificially blurred images, which was superior to the conventional deconvolution method. The knife-edge tests at distances at 200 to 1000 µm from the imaging probe showed an approximately 1.2 times improvement in lateral resolution. In addition, through qualitative evaluation, it was found that the image quality of both ex-vivo and in-vivo tissue images was improved with clear structure and less noise.
Conclusions: This study showed the ability of the 1-D deconvolution network to improve the image quality of OCT images with variable lateral resolution. We were able to train the network with a small amount of data by constraining the network in 1-D. The quantitative evaluation showed better results than conventional deconvolution methods for various amount of blurring. Qualitative evaluation showed analogous results with quantitative results. This simple yet powerful image restoration method provides improved lateral resolution and suppresses background noise, making it applicable to a variety of OCT imaging applications.
Keywords: deconvolution; deep learning (DL); laparoscopy; optical coherence tomography (OCT).
© 2022 Wiley Periodicals LLC.