Abstract
Optical Coherence Tomography (OCT) is a rapidly developing technology for biological tissue imaging. Recently, a great amount of research has been carried out in the denoising and segmentation of retinal OCT images and volumes, but few studies were taken in 3D OCT Angiography (OCTA) for brain microvasculature due to the lack of groundtruth and more challenging imaging conditions, such as the heterogeneous medium. Brain OCTA B-scans (slice of volume) are usually noisier than retinal ones and suffer from motion artifacts that severely degrade the image quality. To address these issues, we propose a novel self-supervised learning method for 3D OCTA volume denoising. Specifically, the center B-scan of each input volume is masked black, and the network is trained to recover it. The only supervision signal the network receives is the original center B-scan itself. Besides, observing the big difference between B-scans with or without motion artifacts, we only select the sample volumes whose center B-scans are unaffected by motion artifacts to avoid introducing bias in training. Such a training strategy can also enable the network against motion artifacts. Experiments show that without any groundtruth training data, our denoising method can effectively deal with both noise and motion artifacts in OCTA volumes.