TY - JOUR
T1 - A Lightweight Swin Transformer-based pipeline for Optical Coherence Tomography Image Denoising in Skin Application
AU - Liao, Jinpeng
AU - Li, Chunhui
AU - Huang, Zhihong
N1 - Publisher Copyright:
© 2023 by the authors.
Funding: This research received no external funding
PY - 2023/4/19
Y1 - 2023/4/19
N2 - Optical coherence tomography (OCT) has attracted attention in dermatology applications for skin disease characterization and diagnosis because it provides high-resolution (<10 μm) of tissue non-invasively with high imaging speed (2–8 s). However, the quality of OCT images can be significantly degraded by speckle noise, which results from light waves scattering in multiple directions. This noise can hinder the accuracy of disease diagnosis, and the conventional frame averaging method requires multiple repeated (e.g., four to six) scans, which is time consuming and introduces motion artifacts. To overcome these limitations, we proposed a lightweight U-shape Swin (LUSwin) transformer-based denoising pipeline to recover high-quality OCT images from the noisy OCT images by utilizing a fast one-repeated OCT scan. In terms of the peak signal-to-noise-ratio (PSNR) performance, the results reveal that the denoised images from the LUSwin transformer (26.92) are of a higher quality than the four-repeated frame-averaging method (26.19). Compared to the state-of-the-art networks in image denoising, the proposed LUSwin transformer has the smallest floating points operation (3.9299 G) and has the second highest PSNR results, only 0.02 lower than the Swin-UNet, which has the highest PSNR results (26.94). This study demonstrates that the transformer model has the capacity to denoise the noisy OCT image from a fast one-repeated OCT scan.
AB - Optical coherence tomography (OCT) has attracted attention in dermatology applications for skin disease characterization and diagnosis because it provides high-resolution (<10 μm) of tissue non-invasively with high imaging speed (2–8 s). However, the quality of OCT images can be significantly degraded by speckle noise, which results from light waves scattering in multiple directions. This noise can hinder the accuracy of disease diagnosis, and the conventional frame averaging method requires multiple repeated (e.g., four to six) scans, which is time consuming and introduces motion artifacts. To overcome these limitations, we proposed a lightweight U-shape Swin (LUSwin) transformer-based denoising pipeline to recover high-quality OCT images from the noisy OCT images by utilizing a fast one-repeated OCT scan. In terms of the peak signal-to-noise-ratio (PSNR) performance, the results reveal that the denoised images from the LUSwin transformer (26.92) are of a higher quality than the four-repeated frame-averaging method (26.19). Compared to the state-of-the-art networks in image denoising, the proposed LUSwin transformer has the smallest floating points operation (3.9299 G) and has the second highest PSNR results, only 0.02 lower than the Swin-UNet, which has the highest PSNR results (26.94). This study demonstrates that the transformer model has the capacity to denoise the noisy OCT image from a fast one-repeated OCT scan.
KW - optical coherence tomography (OCT)
KW - image denoising
KW - deep learning
UR - http://www.scopus.com/inward/record.url?scp=85153705408&partnerID=8YFLogxK
U2 - 10.3390/photonics10040468
DO - 10.3390/photonics10040468
M3 - Article
SN - 2304-6732
VL - 10
JO - Tomography Image Denoising in Skin Application
JF - Tomography Image Denoising in Skin Application
IS - 4
M1 - 468
ER -