Introduction

The research team, led by Chaoliang Chen from Southeast University, compared the performances of amplitude cross-correlation and phase-correlation for calculating translational and rotational offsets in the Fourier-Mellin Transform (FMT) algorithm for optical coherence tomographic angiography (OCTA) images alignment. Their findings revealed that a dual-cross-correlation-based translation and rotation registration (DCCTRR) method significantly reduced the required overlap rate for OCT/OCTA images registration, while maintaining registration accuracy and achieving higher robustness to image background noise. The results are published in Chinese Optics Letters Vol.22 No.7 (2024) (Y. Pu, C Chen, A comparison of cross-correlation based and phase-correlation based image registration algorithms for optical coherence tomographic angiography, 2024, 22(7), 071101) and are selected as the cover paper for that issue.

Research Background

Wide-filed OCT/OCTA imaging offers significant clinical benefits by providing doctors comprehensive lesion information for disease diagnosis. However, current OCT imaging field is limited, and a common way for wide-field imaging is to performing image registration algorithms to stitch multiple overlapping images into a wide-filed image. However, one of the most widely used method, Fourier-Mellin Transform algorithm, requires significant overlap rate for image stitching, which leads to low-efficiency scanning, long acquisition time, and potential motion artifacts, limiting the clinical applications.

Innovative Work

This team proposed DCCTRR for high-accurate, high-efficiency image alignment and large-scale scanning, and compared the performance to traditional FMT method. The results demonstrate that the DCCTRR has the advantage of requiring smaller overlap rate for image stitching and higher robustness to background noise over FMT. The workflow chart of the DCCTRR method is illustrated in Fig.1.

Fig. 1. Workflow of the dual-cross-correlation based translation and rotation registration (DCCTRR) algorithm.

Fig.2. Results of the comparison between DCCTRR and FMT methods. In coefficient map, the horizontal coordinate represents the position of the registered image relative to the reference image, and the vertical coordinate represents the correlation coefficient value at that corresponding position. (a) The normalized cross-correlation coefficient map. (b) The phase-correlation coefficient map. (c) The plot of normalized SNRs of the correlation coefficient map versus different overlap rates. (d) The plot of normalized registration accuracies versus different overlap rates. (e)-(i) Merged images of five representative overlap rates with DCCTRR (the first and the last are the images with the minimum and the maximum overlap rates, respectively). (j)-(n) The merged RGB images of (e)-(i) respectively. (o)-(s) Merged images of the same five pairs of images as (e)–(i) with FMT method. (t)-(x) The merged RGB images of (o) – (s) respectively.

The phantom experiments used fabric fiber to mimic biological microvasculature. From the experimental results in Fig.2, it can be observed that the DCCTRR method exhibits a slower decline in normalized registration accuracy as the overlap area ratio decreases, demonstrating its superior registration performance. As shown in Fig. 2 (i) and (s), compared to the FMT method which requires at least 61.86% overlap for successful registration, the DCCTRR method only needs 12.58% overlap to achieve successful registration, intuitively demonstrating the superior performance of DCCTRR in registering images with low overlap ratios. The team also attempted to add progressively increased Gaussian noise to two images to assess the robustness of the DCCTRR and FMT methods to background noise. The results demonstrated the superiority of DCCTRR over FMT.

Fig.3. The merged large-scale OCTA images with 3x3 OCTA en face images. (a) The left hand of a healthy volunteer (the marked region by a solid line was scanned). (b) The merged large-scale image obtained by DCCTRR.

Furthermore, in vivo experiments were implemented to validate the performance of the proposed DCCTRR, and the results are shown in Fig. 3 which demonstrated the that the proposed method has the capability of stitching images with an overlap rate of ~20% for wide-field OCTA imaging and could significantly improve the scanning efficiency in clinical applications.

Conclusions and Prospects

The research team investigated the effectiveness of a double cross-correlation-based translation and rotation registration (DCCTRR) method that utilizes amplitude cross-correlation instead of phase correlation to calculate translational and directional offsets within the workflow of the FMT algorithm. The performance of the DCCTRR method was compared to FMT method in OCT/OCTA image stitching. Both phantom and in vivo experimental results demonstrated that the DCCTRR method requires a smaller overlap rate than the FMT method to achieve comparable registration accuracy, and the DCCTRR method is more robust to background noise.

In the future work, the research team plans to improve the computing efficiency of the proposed method by applying parallel computing technologies (such as GPU and FPGA) to achieve near real-time processing speeds. Additionally, the team next step will explore performing the proposed DCCTRR for aligning the images obtained by robotic arms assisted wide-field scanning, which could further improve scanning speed and imaging speed. These works may be able to benefit the applications of OCT/OCTA imaging in clinics, and also contribute to the development of robotic arms assisted OCT/OCTA imaging.