Automated Artifacts and Noise Removal from Optical Coherence Tomography Images Using Deep Learning Technique

摘要

Optical Coherence Tomography (OCT) is a popular non-invasive clinical tool for the diagnosis of ocular diseases that provides micron-scale images of ocular pathology in vivo and in real-time. The cross-sectional OCT B-scan of Temporal-SuperiorNasal-Inferior-Temporal (TSNIT) peripapillary retinal profile is widely used to diagnose and monitor glaucoma. However, raw OCT images can be marred by noise and artifacts, especially vitreoretinal interface opacity: this can lead to segmentation error, misinterpretation of retinal thickness measurements and possibly inappropriate glaucoma management. In this study, we designed and trained a U-Net model on OCT B-scans with artifacts, and their corresponding artifact-free B-scans’. The U-Net was able to remove the artifacts successfully with better performance in terms of PSNR and SSIM values. The SNR of the OCT scans with speckle noise associated with artifacts has also been improved. To the best of our knowledge, this is the first study where automated vitreous opacity artifact removal has been applied to the TSNIT profile. The performance of the U-net model on measures such as PSNR, SSIM, MAE, and MSE is compared with the state-of-the-art image denoising models. It is observed that the proposed U-Net model performs better as compared to the other models on both parametric and visual evaluations. In the future, this U-Net model could be used to solve automatic retinal layer segmentation errors and assist clinicians in interpreting OCT images in glaucoma diagnosis and monitoring.