Speckle noise reduction in swept-source optical coherence tomography by retinal image registration

The averaging of multiple B-Scans is an effective method of reducing speckle noise in Swept-Source Optical Coherence Tomography (SS-OCT) and obtaining clear structural information. However, physiological characteristics such as eye tremor, drift, micro-saccade, and the optical structure of an SS-OCT system cause geometric transformation between images, resulting in poor multi-frame averaging. In this paper, we propose a registration algorithm based on the combination of gray distribution information and target geometric information. This method extracts the region of interest containing target information using the average gray distribution of an image, and corrects the transformation of the image with the collective effect of the phase correlation algorithm and the gray projection algorithm based on the fitting of the curve of its segments. Then, the process is repeated with the upper boundary of the retinal image fitted as the feature points to determine the optimal rotation parameters. The translation parameters are re-estimated again to achieve the rigid registration of the image. Finally, a one-to-one mapping method of axial scanning is used to achieve the non-rigid registration of the image with the energy function as the constraint. Experiments on live rabbit eyes show that the averaged image has clear boundaries, enhanced structural information, and its signal-to-noise and contrast-to-noise ratios are more than doubled their previous values on average. The algorithm is suitable for the registration of B-Scan images with strong speckle noise and can meet the averaging needs of many types of OCT systems. It has high robustness and image registration accuracy.