Optical coherence tomography angiography (OCTA), a representative functional extension of OCT, has been widely utilized in ophthalmology clinics for its capabilities of in vivo label-free vascular imaging to monitor and diagnose retinal disorders such as glaucoma, diabetic retinopathy, and age-related macular degeneration [1]. Methods to minimize motion artifacts in a clinical environment are highly demanded, as patients with retinal diseases tend to have poor fixation and excessive eye movements. Previously, we reported multi-volume 3D registration and averaging processing to mitigate the motion artifacts of OCTA images [2]. A complementary solution to diminish motion artifacts is to reduce the acquisition time with a faster A-scan rate OCT systems. Fourier-domain mode-locking (FDML) is an advanced swept-source laser technique capable of achieving multi-MHz sweep rates, which allows ultrahigh-speed retinal image acquisition [3]. Based on its superiority in speed, MHz-rate ophthalmic imaging with FDML has a high potential to be implemented into clinical practices.

However, several limitations of the FDML laser remain to be addressed for the application of OCTA imaging. One of the current drawbacks of FDML is its relatively low system sensitivity, resulting in limited visibility of small capillary vessels. Although the signal-to-noise ratio (SNR) can be improved by acquiring and averaging multiple repeated B-scans (BM scans) [3,4] or multi-volume data [5], these solutions come at the cost of longer acquisition time and/or imposing a higher computational load for image registration. In a recent report, Pfeiffer et al. suggested coherent averaging of multiple A-scans as a simple and effective way to increase the SNR of FDML-based MHz OCT systems [6]. While this is encouraging, especially since oversampling in the fast-scan direction is unavoidable in many FDML systems due to the current mechanical limitation of scanners [3,4], the effect of coherent A-scan averaging on retinal OCT imaging has not been thoroughly investigated. Additionally, when applying coherent A-scan averaging to FDML lasers, the impact of fiber dispersion in the buffering stage should be carefully considered, as it induces phase variation between adjacent A-scans. Although the impact of fiber dispersion for longer wavelengths such as 1300 nm may be negligible, as the authors of Ref. [6] described, fiber dispersion at 1060 nm could be significant, a wavelength that is commonly used for a retinal scan. Hence, minimizing fiber dispersion effects on the buffering stage may improve phase alignment among A-scans and enhance the quality of 1060 nm FDML OCTA systems.

In this Letter, we demonstrate phase-corrected coherent buffer averaging as an effective approach to improve SNR of retinal OCTA in FDML-based systems. A numerical spectral resampling algorithm is proposed to reduce phase differences among buffers in the FDML source, with results indicating significant improvement in phase alignment.

To apply coherent averaging, the signals need to be properly aligned in phase before averaging [7]. In the case of coherent A-scan averaging with a FDML laser, variability in the spectral characteristics among the buffered spectra is one of the main contributing factors for phase mismatch. Therefore, we first investigated the effect of fiber dispersion in buffering stages of the laser. Spectral interferograms from a stationary mirror were acquired without scanning (Fig. 1). In this study, we employed a 1.6 MHz FDML laser (NG-FDML-1060-4B-FA, OptoRes GmbH, Germany) centered at 1060 nm with a 75 nm tuning range. The FDML laser increases the A-scan rate by optical buffering to create three successive time-delayed copies of the original sweep using long optical fiber spools. We will refer to each of the buffer signals (i.e., four successive A-scans) from one cluster as buffer 1, buffer 2, buffer 3, and buffer 4, based on the time delay from the trigger signal. The representative four buffer spectra, each consisting of 1152 sampling points based on the ${\sim}{1.8}\;{\rm{GHz}}$ sampling clock signal generated from a phase-locked loop of the source, were digitized from a single trigger starting point. In this study, each buffer set is acquired with a separate trigger signal from the laser source. The amplitude differences are noticeable among the four buffer sweeps, mainly coming from imperfections in the splitting ratio of the fiber couplers in the buffering stage. Spectral mismatch in the interference signals can be observed at specific parts of the interferograms, which leads to phase distortion in the A-scans originating from different buffers.

 

Fig. 1. Spectral interference overlays of four buffer signals.

Download Full Size | PDF

The proposed numerical buffer-dependent spectral resampling algorithm converts the acquired interferograms into wavenumber liner (k-linear) representations followed by an optimization step to minimize the spectral differences between buffers. In the first step, a phase linearization algorithm [8] is used to create timing resampling vectors that rescale the spectra into a k-linear representation as shown in Fig. 2(A). However, as FDML data consist of multiple buffers (i.e., four), each with slightly different spectral characteristics, the timing vector for one buffer signal may not be optimal for the other buffered signals. To further investigate this effect, timing vectors generated using buffer 1 are applied to other buffers in Fig. 2(B), where slight differences in the unwrapped phase slope can be observed. To address this buffer dependency issue, a polynomial timing delay term is added to the timing vectors generated from buffer 1, similar to the approach used in Ref. [9], as follows:

 

Fig. 2. Phase characteristics of each buffer interferogram in (A), (B) wavenumber linearization and (C), (D) phase optimization step.

Download Full Size | PDF

Here, ${T_1}$ is the timing vector of buffer 1, ${T_N}$ is the timing vector of the target buffer (i.e., $N = {{2}}$, 3, 4), $t$ is the timing index, and ${a_{i,N}}$ is the $i$th order polynomial fitting parameter for buffer $N$. Then, ${T_N}$ is optimized iteratively by minimizing the fringe washout due to the relative spectral shift between the target buffer interferogram and the buffer 1 interferogram after spectral resampling. Fringe washout was estimated based on the power decrease caused by the signal cancellation and defined as

To evaluate phase slope differences among buffers, the unwrapped phase slope of each buffer is subtracted from buffer 1. With only the phase linearization step, there are noticeable differences in the phase among buffers [Fig. 2(C)]. As expected, the amount of delay in the fiber spool in the buffering stage affects the amount of phase mismatch, with buffer 4 having the largest phase difference compared to buffer 1. With numerical phase correction, as shown in Fig. 2(D), the phase slopes of the buffers show smaller differences, indicating the spectral differences among buffers are significantly reduced.

To evaluate the effects of phase optimization on A-scans, Fourier transformed signals from a single reflector placed at different depth positions are presented. Figure 3(A) shows representative A-scan plots of four buffers without phase optimization, indicating variations among buffers that can contribute to point-spread-function broadening when averaged. Figure 3(B) is the A-scan profiles of four buffers after phase optimization and amplitude correction, showing the A-scan signals among buffers become nearly identical. As Fig. 3(C) indicates, the amplitude biases in the source are independent of depth, and therefore, a simple amplitude ratio can be applied to the acquired interferograms for correcting the amplitude differences among buffers. Figure 3(D) represents A-scan phase differences of each buffer A-scan compared to buffer 1 within the same cluster. A total of 100 buffer sets (i.e., 400 A-scans) was acquired at each position. Without phase correction, shown in solid lines, phase alignment incrementally worsens as the depth increases, with the buffer 4 signals having the most significant difference of 0.59 radians on average. Phase correction, indicated by dashed lines, greatly improved the phase alignment among A-scans to an average of 0.05 radians.

 

Fig. 3. Intensity profiles of a mirror signal (A) without and (B) with phase optimization and amplitude correction. (C) A-scan intensity ratio and (D) A-scan phase differences compared to the buffer 1.

Download Full Size | PDF

To investigate the contrast enhancement from phase-corrected buffer averaging, in vivo retinal OCTA images were acquired from a healthy volunteer in accordance with the human study protocol approved by the institutional review board at UBC. The FDML system was built based on a typical swept-source retinal OCT configuration and has axial and lateral resolutions of 9 µm and 13 µm, respectively. The scanning beam is delivered by a two-axis galvanometer scanner (6210H, Cambridge Technology, USA) with optical power on the cornea configured to be below 1.5 mW. The scan protocols employed in this study are summarized in Table 1. To minimize the fly-back portion and the acquisition time, we employed a stepping bi-directional scanning protocol with a two-step interval [10]. In particular, the 2BM slow scan protocol is intended to take full advantage of coherent averaging along the A-scan direction, as the pixel resolution is more than four times smaller than the system lateral resolution even under the largest field of view (FOV) we tested (i.e., $6 \times 6\,\,{\rm{mm}}$). This ensures that effective pixel resolution is sustained even after the A-scan averaging. Baumann et al. suggested coherent averaging can more effectively reduce noise than magnitude averaging if the signal phase is properly aligned [7]. In this study, we coherently averaged up to 4A-scans belonging to the same cluster, as the phase among buffers is highly stable without trigger jitter. Figure 4 shows the OCT B-scan and corresponding OCTA B-scan acquired using the 2BM scan protocol. OCTA was obtained by calculating variance of complex OCT signals across repeated B-scans, which provided microvascular contrast superior to amplitude-based calculation [11]. Before buffer averaging, the OCT image has a high background noise level, and tissue structures in the inner retina and choroid can hardly be resolved. In the OCTA B-scan, the high noise floor makes it impossible to distinguish small capillaries from the noise. After averaging 4A-scans, there is a noticeable SNR improvement in both OCT and OCTA images. The OCT B-scan shows more distinct retinal layer boundaries and higher contrast in the choroidal region. In the OCTA B-scan, noise suppression and signal enhancement from averaging reveals small capillary vessels as indicated by the yellow arrows. The OCTA image quality can be further improved as correcting spectral shifts caused by trigger jitter among different clusters by adopting the previously reported method [10] with additional calibration steps.

Table 1. Scanning Protocols

View Table

 

Fig. 4. OCT and complex variance OCTA B-scans of human macula.

Download Full Size | PDF

En face projection of OCTA using different numbers of A-scan averaging is presented in Fig. 5. The superficial vascular complex (SVC) was separated from the deep vascular complex (DVC) using 3D graph-cut-based segmentation [12]. Here, image contrast was adjusted based on the individual OCTA volume. Without A-scan averaging, the 2BM OCTA can clearly distinguish features only in the SVC, and the DVC is contaminated with a high level of noise. A 2BM scan with 2A- and 4A-scan coherent averaging effectively suppressed the noise floor and enhanced the visibility of features in both the SVC and DVC. In practice, one can select the number of A-scans to average based on their system requirements such as imaging speed, sampling density, FOV, and SNR. For instance, one can simultaneously acquire a large FOV with 2A-scan averaging and a small FOV with 4A-scan averaging for higher sensitivity without any hardware modification.

 

Fig. 5. En face OCTA projection of the human macula, comparing different A-scan averaging strategies. FOV: $6 \times 6\,\,{\rm{mm}}$. Scale: 1 mm.

Download Full Size | PDF

Last, we compared the 2BM slow scan protocol and the 4BM fast scan protocol under comparable scanning conditions. Increasing the number of repeated BM scans using the 4BM scan protocol at a high frame rate improved the vessel continuity and signal intensity at the cost of longer volume acquisition time as shown in Fig. 6. However, the performance of this scanning protocol is highly dependent on the mechanical response and repeatability of the scanner, making it difficult to implement in a clinical setting where high system stability is required. On the other hand, the 2BM scan with four buffer coherent averaging effectively suppressed the noise floor and enhanced the visibility of both SVC and DVC. A side-by-side comparison of 4BM and 2BM OCTA protocols shows that 2BM with A-scan averaging provides better OCTA contrast. Combining coherent buffer averaging with multi-volume 3D registration can further improve the image contrast by effectively reducing the speckle noise and increasing vessel connectivity.

 

Fig. 6. En face OCTA projection of the human macula, comparing different acquisition protocols. FOV: $3.5 \times 3.5\,\,{\rm{mm}}$. Scale: 1 mm.

Download Full Size | PDF

In summary, we introduced phase-corrected buffer averaging to improve the SNR of retinal OCTA in FDML-based systems. We presented a numerical phase correction method to mitigate spectral differences among buffers of the FDML and effectively increase the SNR through coherent averaging. The proposed approach to compensate for the reduced sensitivity in ultrahigh-speed retinal scanning has unveiled the potential of FDML-based systems to be used for multiscale imaging, such as the multiscale adaptive optics systems we previously demonstrated [13]. Future studies will explore effective scanning protocols for multiscale FDML imaging with various FOVs, resolutions, and sensitivities.