SELF-SUPERVISED DENOISING OF OPTICAL COHERENCE TOMOGRAPHY WITH INTER-FRAME REPRESENTATION

Spectral-domain optical coherence tomography (SD-OCT) is a high-speed ocular imaging technology that is commonly employed in eye examinations to visualize the back structures of the eyes. OCT volume containing a sequence of cross-sectional images can be captured in seconds. However, the low signal-to-noise ratio (SNR) prevents accurate result interpretation. To obtain a high SNR OCT volume, numerous images must be averaged at each imaging depth, which is time-consuming. Subjects, especially children, who have short attention spans, may significantly hinder the data collection procedure. Most of the current algorithms focus on single-frame processing without using inter-frame information. Here we developed a lightweight 3D-UNet with a self-supervised strategy to denoise the low SNR OCT volume. This method does not require noisy-clean pairs and can be accomplished by simply measuring a volume containing multiple OCT images. The proposed method improves image quality with structural details preserved and achieves state-of-the-art performance on real OCT datasets.