Open Access
Add to BibSonomy
Abstract
Dynamic OCT angiography (OCTA) is an attractive approach for monitoring stimulus-evoked hemodynamics; however, a 4D (3D space and time) dataset requires a long acquisition time and has a large data size, thereby posing a great challenge to data processing. This study proposed a GPU-based real-time data processing pipeline for dynamic inverse SNR-decorrelation OCTA (ID-OCTA), offering a measured line-process rate of 133 kHz for displaying OCT and OCTA cross-sections in real time. Real-time processing enabled automatic optimization of angiogram quality, which improved the vessel SNR, contrast-to-noise ratio, and connectivity by 14.37, 14.08, and 9.76%, respectively. Furthermore, motion-contrast 4D angiographic imaging of stimulus-evoked hemodynamics was achieved within a single trail in the mouse retina. Consequently, a flicker light stimulus evoked an apparent dilation of the retinal arterioles and venules and an elevation of the decorrelation value in the retinal plexuses. Therefore, GPU ID-OCTA enables real-time and high-quality angiographic imaging and is particularly suitable for hemodynamic studies.
© 2022 Optica Publishing Group under the terms of the Optica Open Access Publishing Agreement
1. Introduction
Motion-contrast OCT angiography (OCTA) allows fast and safe 3D visualization of blood perfusion to the capillary level [1–4], and decorrelation-based OCTA has been widely used in clinical and preclinical systems [1–8]. Owing to the normalization operation, decorrelation generates apparent false-positive blood flow signals (high decorrelation induced by random noise) in the low signal-to-noise ratio (SNR) area. Thus, to suppress false-positive signals, an SNR-adaptive flow mask has been generated based on an asymptotic linear relation between the inverse SNR (iSNR) and decorrelation (ID), and an ID-OCTA was developed [9]. In addition, the decorrelation signal is closely related to flow dynamics [10–16]. Thus, dynamic ID-OCTA is an attractive approach for monitoring stimulus-evoked hemodynamic activities.
However, dynamic imaging poses a significant challenge for data processing. Typically, dynamic OCTA requires a considerable time (tens and even hundreds of seconds) to scan a certain 3D volume repeatedly and record the hemodynamics over a certain time. During such a long-time interval, both defocusing and bulk motion may result in failure of data acquisition, which can be reduced through the proper adjustment of OCTA system and improvement of subjects’ comfort [17,18]. Real-time feedback of image quality can aid in adjusting the system and acquiring usable data within a limited time. Practically, in decorrelation estimation, the limited kernel size results in a considerable intermediate region with a mixed distribution of static and dynamic voxels in the ID space [19]. Consequently, this poses a great challenge to determining an appropriate ID threshold for reserving dynamic voxels and removing static voxels. Automatic optimization of the ID threshold can be achieved based on image quality feedback. Thus, real-time data processing is desired to generate feedback of vascular network visibility and accordingly guide the optimization of ID-OCTA data acquisition and processing.
Similar to most OCTA techniques, ID-OCTA is computationally intensive, involving a fast Fourier transform (FFT), spatial alignment, layer segmentation, and ID flow mask generation. Further, a single 3D OCTA dataset usually requires more than tens of seconds in the conventional CPU-based ID-OCTA processing framework [9]. Since the graphics processing units (GPU) have efficient parallel structure for large volume data computation, GPU has been used to realize fast OCT with megahertz processing line rate [20], achieve the real-time layer segmentation combining deep learning with labeled ground truth [21], and increase the OCTA processing speed [17,18,22,23]. Inspired by these literatures, the GPU is promising for accelerating ID-OCTA processing speed via parallel computation and allowing real-time angiographic imaging.
In this study, a GPU-based real-time data processing pipeline was proposed for dynamic ID-OCTA imaging. The time consumption and image quality of the proposed GPU ID-OCTA were quantitatively analyzed. Further, 4D angiographic imaging of flicker light stimulus (FLS)-evoked hemodynamics was demonstrated with GPU ID-OCTA in the mouse retina in vivo. Furthermore, the implications and limitations of this study are discussed.
2. Materials and methods
2.1 Animal preparation and system setup
All experiments were conducted on 12-week-old male C57BL/6J mice (N = 5) in a dark room. Prior to performing the experiments, the mice were adapted to dark for 1.5 h and then anesthetized and fixed in a custom-made animal holder to minimize the influence of both heart beating and breathing, the pupil was dilated and the cornea was moisturized with sodium hyaluronate solution to minimize the evaporation of tear film over the entire experiments, and the body temperature was maintained using a heating pad [24]. All animals were obtained from the Zhejiang Medical Science Institute and all experimental procedures performed in this study were approved by the Animal Care and Use Committee of the Zhejiang University.
Retinal FLS were generated using a green light-emitting diode (520 nm wavelength). Each FLS pulse had a 50 ms pulse width, and the pulse trains were delivered at a repetition rate of 10 Hz. Further, the laser power of FLS detected on the pupil is ∼0.2 mW, and the combined power of FLS and OCT light on the pupil is ∼1.0 mW, which was within the American National Standards Institute safety limit [25]. Moreover, each trial comprised a 30 s pre-stimulus period, followed by a 30 s FLS (i.e., a train of 300 pulses).
FLS-evoked retinal hemodynamic responses were recorded using a lab-built spectral domain (SD) OCTA system. The light source was a superluminescent diode with a central wavelength of 840 nm and a spectral bandwidth of 100 nm, yielding a measured axial resolution of ∼4 µm (full width at half maximum) in air. The probe beam was focused on the retina with a ∼0.5 µm beam size on the cornea, providing a lateral resolution of ∼10 µm in the retina. The spectral interference fringes were recorded with a line-scan CMOS camera (e2v, UK) with 2048 active pixels and 120 kHz line-scan rate. Further, each OCTA volume contained 768 B-scans repeated three times at 256 tomographic positions (slow-scan, y-direction) with each B-scan composed of 2048 × 256 (z × x) voxels, corresponding to a total acquisition time of ∼2 s. In addition, repeated volumetric OCTA scans were centered on the optical nerve head with a field of view (FOV) of 2 mm × 2 mm (x-y) and performed at a time interval of 6 s to record the time course of FLS-evoked responses. The FLS was synchronized with the OCTA imaging via a custom-designed trigger module.
2.2 GPU accelerated ID-OCTA
A custom GPU parallel processing program was developed using CUDA version 11.0, on a 64-bit Windows 10 operating system, equipped with NVIDIA GeForce RTX 2080 Ti GPU, Intel Core i7-9700k 3.60 GHz CPU, and 32.0GB RAM. The proposed GPU ID-OCTA process is depicted in the flowchart shown in Fig. 1. A total of 12 B-frames of raw data (four tomographic positions) were batched as a processing unit to improve the efficiency of data transfer and GPU computation [22,26]. On acquiring a new batch, it was transferred from the data acquisition (DAQ) buffer to the GPU device memory and sequentially processed by SD-OCT and ID-OCTA kernels.
Fig. 1. Processing flow chart of the proposed GPU ID-OCTA. DAQ, OCT and OCTA buffer are located in CPU memory (green panel). OCT kernel, ID-OCTA kernel and ID threshold optimization were executed in GPU memory (yellow panel). In ID-OCTA kernel, batch matrices of intensity, iSNR, decorrelation and layer location were temporarily stored and stacked up as 3D matrices when all batches were acquired and processed. The batch processing timeline of data transfer, OCT kernel and ID-OCTA kernel is provided in the red panel. Blue panel: Procedures of ID-OCTA kernel. Each volume complex data contains 3 $\mathrm{\ast }$Y B-scans repeated 3 times at Y tomographic locations and each B-scan is composed of X A-lines with the length of Z pixels. (a) Calculate the OCT intensity signal from the input aligned 3D OCT complex data (named as Complex1). Take the average value of 3 repeated scans following modulus operation as the Intensity matrix. (b) Take the average value of 3 repeated scans from the intensity matrix and segment layers. (c) Obtain iSNR matrix from intensity matrix. (d) Move 1 page backward for each page in Complex1 to generate Complex2. The first and last page of Complex1 are outlined with red and green, respectively. (e) Perform correlation operation to Complex1 and Complex2 in 3rd dimension. Discard invalid pages (painted with yellow), and take the average value of every 2 pages of the rest matrix as correlation matrix. (f) Obtain decorrelation matrix. (g) Generate ID-mask with original ID threshold ${\gamma _0}$.
In the SD-OCT kernel, tasks including k interpolation, dispersion compensation, FFT, and correlation-based alignment are executed sequentially [7]. Parallel computations are applied to each task. Whereas, in the ID-OCTA kernel (blue inset in Fig. 1), the intensity, iSNR, and decorrelation matrices were computed. Subsequently, the ID-OCTA signal were extracted by ID-mask, and the retinal layers were segmented. Further, the structural and angiographic cross-sections were produced and transferred to the CPU buffer for real-time display. On completion of the acquisition and processing of an entire 3D scan (64 batches), a Gaussian filter (3×3×3 kernel size) was applied to the 3D OCTA dataset to suppress the background noise, and en face OCTA images were generated immediately with maximum value projection (MVP) from the retinal slabs segmented by layer location.
In this study, a GPU-based layer segmentation method was developed. The inner limiting membrane (ILM) and the retinal pigment epithelium (RPE) was automatically segmented via the application of an intensity variation method [27] to all A-lines of a 3D data in parallel. Accordingly, the inner plexiform layer (IPL), inner nuclear layer (INL), and outer plexiform layer (OPL) were estimated, and five OCTA slabs were produced: 1) superficial vascular plexus (SVP, from ILM to IPL), 2) intermediate capillary plexus (ICP, from IPL to INL), 3) deep capillary plexus (DCP, from INL to OPL), 4) retina (RE, from ILM to OPL), and 5) choroid (CH, below RPE).
In ID-OCTA, to extract the red blood cell (RBC) motion contrast, complex decorrelation was computed between the repeated B-frames with a local spatio-temporal kernel [9], generating a raw decorrelation mapping. In the mapping, the decorrelation D of the static voxels (e.g., surrounding tissues) has an asymptotic linear relationship with the iSNR [9]:
where $iSNR = \frac{{{s^2}}}{I}$, ${s^2}$ denotes the system noise level, I denotes the intensity of OCT signals, and a.s. denotes convergence with a probability of one. In addition, static voxels have a considerable intermediate region ${R_i}$ with dynamic voxels (e.g., blood flow) in the ID space. Region ${R_i}$ can be demarcated with two boundaries: the higher limit of static voxels ${D_1}$ (blue line in Fig. 2(a)) and the lower limit of dynamic voxels ${D_2}$ (red line in Fig. 2(a)) [19].Fig. 2. ID space and the angiograms quality evaluation metrics. (a) Illustration of ID space. The blue and red lines represent the boundary lines corresponding to ${D_1}$ and ${D_2}$ respectively. Point F is their intersection point. ${R_d}$, ${R_i}$ and ${R_s}$ are the dynamic region, intermediate region and static region. The angle $\gamma $ between the threshold line ${D_i}$ and the decorrelation axis indicates the ID threshold. (b) The en face RE OCTA images and the $QI$ values corresponding to different $\gamma $ from same data. (c) Intensity gradient curves of ${L_1}$-${L_4}$ in (b). The horizontal axis is the pixel position, and the longitudinal axis means the intensity gradient. (d) The changes of normalized $VD$, ${C_{RMS}}$ and $QI$ over $\gamma $ corresponding to the data in (b).
Referring to Eq. (1), in addition to the dynamic flow, static tissues with low SNR (e.g., deep regions) also present high decorrelation values, resulting in severe false positive flow signals. Thus, to reduce the decorrelation artifacts, an ID flow mask was generated as follows: 1) projecting all voxels from the XYZ space into the ID space, 2) designing an ID threshold based on the features of both iSNR and decorrelation, 3) assigning the dynamic voxels as one and the remaining as zero in the ID space, and 4) projecting all voxels from the ID space into XYZ space. The ID-mask was overlaid on the raw decorrelation mapping to generate the angiograms. During batch processing, the higher limit of the static voxels ${D_1}$ was used as the initial ID threshold to obtain a quick glance at the OCTA cross-section. However, a direct cutoff along the higher limit of static voxel ${D_1}$ excluded the considerable dynamic signals of ${R_i}$. An optimal ID threshold should achieve a trade-off between the reservation of dynamic voxels and the removal of static voxels for high-quality OCT angiograms.
2.3 Automatic optimization of ID-OCTA
Owing to the fast processing in GPU ID-OCTA, auto-optimization of the ID threshold can be achieved by maximizing the quality of the en face angiogram. As shown in Fig. 2(a), passing through the intersection F of lines ${D_1}$ and ${D_2}$, several threshold lines ${D_i}$ (orange line in Fig. 2(a)) can be selected between the ${D_1}$ and ${D_2}$. The angle between the threshold line and the decorrelation axis is denoted as $\gamma $. Different $\gamma $ values corresponded to a change in vascular network visibility in the en face angiogram (Fig. 2(b)). To evaluate the en face angiogram quality when ${D_i}$ traversed the intermediate region ${R_i}$, a quality index ($QI$) was defined as the product of the vessel density ($VD$) and the global contrast of the OCTA image (root-mean-square contrast, ${C_{RMS}}$):
where VD was computed as: where L denotes the side length of en face OCTA image, and a total of four pairwise perpendicular lines ${L_1}$-${L_4}$ (red dashed lines in Fig. 2(b)) located at $\frac{L}{4}$ and $\frac{{3L}}{4}$ on both sides were selected to calculate ${V_i}$: 1) the intensity gradient was calculated from the en face angiogram along each line and plotted versus the pixel position (X) for each line (Fig. 2(c)); 2) ${V_i}$ was counted as the number of intersections between the horizontal axis and the intensity gradient curve of line ${L_i}$. The en face OCTA image presented a sharp change in brightness at the vessel location, and the frequency of this change considered by ${V_i}$ as a global index of $VD$.The root-mean-square contrast ${C_{RMS}}$ of the en face image (denoted as $A$) was computed as [28,29]
As shown in Fig. 2(d), with increase in the threshold angle $\gamma $, both $VD$ and ${C_{RMS}}$ of the angiogram from the same data were first improved via reserved blood signals; then, $VD$ was saturated because most vessels had been extracted and ${C_{RMS}}$ was impaired by redundant noise. Consequently, the $QI$ is a convex function of angle $\gamma $. The optimal threshold angle ${\gamma _T}$ was defined as the angle at which $QI$ was maximized, and was automatically determined using a ternary search method [30]. Subsequently, according to ${D_1}$ and ${D_2}$, two starting points of the ID threshold were defined: ${\gamma _1}$, ${\gamma _2}$ (${\gamma _1} < {\gamma _2})$. The number of iterations M was empirically set to six. In addition, ${\gamma _1}$ and ${\gamma _2}$ converged to ${\gamma _T}$ in the last iteration. In the algorithm, all the raw materials, such as matrices of iSNR, decorrelation, and layer location, were readily available, and the processes of angiogram generation and $QI$ calculation were implemented in parallel in the GPU for each iteration. Hence, the automatic optimization of the ID-OCTA was highly efficient.
2.4 Metrics for OCTA image quality
To quantitatively evaluate the performance of the auto-optimization of the ID threshold, three additional metrics were used: SNR, contrast-to-noise ratio (CNR), and blood vessel connectivity (CON) [31]. First, the angiograms were binarized based on the Otsu threshold [32], then the background area was defined as the non-perfused region where the value was 0 in binarized images [33], and subsequently skeletonizing morphological operations were performed to obtain the vessel skeleton maps. In the vessel skeleton maps, pixels with more than five consecutive connections (including diagonal connections) were labeled as connected flow pixels. Following preprocessing, the SNR, CNR, and CON were calculated as follows:
3. Results
3.1 Time consumption of GPU-accelerated ID-OCTA
The processing time of a single batch in the proposed GPU ID-OCTA was captured using NVIDIA Visual Profiler. As reported in Table 1, each batch required an average time of 23 ms (equivalent to a 133 kHz line-process rate), of which ∼46% of the time consumption was used for alignment and ∼23% for segmentation. Thus, real-time ID-OCTA processing was achieved in the current 120 kHz line-scan rate system. Following the processing of all batches, an additional 0.12 s was required in the operation of MVP and ID threshold auto-optimization. Compared with the CPU-based ID-OCTA (∼932 ms for the same batch), the proposed GPU pipeline presented a ∼40 times improvement in processing speed.
Table 1. Time consumption comparison of single batch between CPU- and GPU-based ID-OCTA. Each batch has 12 B-frames (repeated 3 times at 4 Y-positions), and 2048 × 256 (Z × X) voxels for each frame. Test was performed on the mouse retinal imaging in vivo. Each test was repeated on 10 subjects.
3.2 In vivo imaging of auto-optimized ID-OCTA
Cross-sectional and en face angiograms were successfully obtained with the proposed GPU ID-OCTA processing pipeline in the mouse retina in vivo. As shown in Fig. 3(a), ILM (line #1) and RPE (line #5) were automatically segmented based on the intensity variation in the depth direction. Further, the retinal volume was flattened according to the RPE layer to avoid the influence of the retinal curvature on rendering. Accordingly, the IPL (line #2), INL (line #3), and OPL (line #4) were estimated using an empirical algorithm [34]. The layer information was directly applied to the corresponding angiogram (Fig. 3(b)), and en face angiograms of the RE, SVP, ICP, DCP, and CH slabs were readily generated (Figs. 3(c)–3 g) with an auto-optimized ID threshold. The black line in Fig. 3(d) (indicated by the red arrow) was due to a misclassification of SVP into ICP, and further improvement of the empirical segmentation algorithm was in need around the optic nerve head.
Fig. 3. Mouse retina imaging with GPU ID-OCTA. Structural (a) and angiographic (b) cross-sections along the red dashed line in (c). The inner limiting membrane (ILM, line #1) and the retinal pigment epithelium (RPE, line #5) were segmented automatically, and the inner plexiform layer (IPL, line #2), the inner nuclear layer (INL, line #3) and the outer plexiform layer (OPL, line #4) were estimated accordingly. Retina was flattened according to RPE. (c-f) En face angiograms of retina (RE, from ILM #1 to OPL #4), superficial vascular plexus (SVP, from ILM #1 to IPL #2), intermediate capillary plexus (ICP, from IPL #2 to INL #3), deeper capillary plexus (DCP, from INL #3 to OPL #4) and choroid (CH, below RPE #5) slabs, respectively. The red arrow in (d) indicates a misclassification of SVP into ICP. The field of view was 2 × 2 mm2 centered on the optical nerve head, and the imaging time was ∼2 s.
The superiority of the auto-optimized ID threshold was further validated. The ground truth of the retinal vasculature was generated by averaging ten angiograms that were acquired sequentially from the same retinal volume with SID-OCTA [19]. SID-OCTA applies a shape mask to the intermediate region ${R_i}$ to identify the dynamic flow [19]. When turning the optimization off, a conserved threshold ${\gamma _1}$ was used, and region ${R_i}$ was completely removed [9]. Compared with the ground truth (Fig. 4(a)), the conserved threshold ${\gamma _1}$ resulted in the overkilling of the flow signals (green pixels in Fig. 4(b) and insert I), which was reserved with the optimized threshold ${\gamma _T}$ at the price of more background noise (purple pixels in Fig. 4(c) and insert II). As an example, a DCP vessel (see the arrow heads in Fig. 4(d)–4(f)) was overkilled by the threshold ${\gamma _1}$, but was well preserved by the optimized threshold ${\gamma _T}$. As shown in Table 2, compared to the conserved threshold ${\gamma _1}$, the optimized threshold resulted in an improved angiogram quality of 14.37% in SNR (p = 0.006), 14.08% in CNR (p = 0.018), and 9.76% in CON (p = 0.044). This suggested that the optimized threshold ${\gamma _T}$ achieved an improved trade-off between reserving flow voxels and removing background voxels in region ${R_i}$.
Fig. 4. Representative ID-OCTA images with optimization off and on. (a) Ground truth of retinal vasculature. It was generated by averaging ten angiograms that were acquired sequentially from the same retinal region with SID-OCTA [19]. (b) Fusion of ground truth (green) and angiogram without optimization (purple). (c) Fusion of ground truth (green) and angiogram with optimization (purple). (d-f) Corresponding cross-sectional angiograms along the dashed lines in (a-c), respectively. (I and II) Enlarged view of the rectangular regions in b and c, respectively. In the 2nd and 3rd columns, the green and purple colors might indicate an overkill of dynamic flow and an underkill of static background, respectively. The red triangles indicate the apparent changes before and after optimization. Averaging over different dataset was only performed in a, and only one dataset was used in b and c.
Table 2. Quantitative comparison between the ID-OCTA images with optimization off and on. A total of 10 RE slabs from different samples were used for statistics. Three metrics were calculated: signal-to-noise ratio (SNR), contrast-to-noise ratio (CNR) and connectivity (CON). Results are presented as mean ± std. Δ means the improvement.
3.3 FLS-evoked hemodynamic response of mouse retina in vivo
Owing to the rapid capture of high-quality angiograms, the proposed GPU ID-OCTA facilitated dynamic angiographic imaging of FLS-evoked hemodynamics in the mouse retina in vivo, as shown in the cross-sectional and en face angiographic video (Visualization 1). Compared with the baseline (Fig. 5(a)–5(c)), the FLS evoked an apparent dilation of the arterioles and venules in the SVP (Fig. 5(d), insert I and II) and an elevation of the decorrelation value in all three slabs (Fig. 5(d)–5(f), insert III-VIII). As plotted in the time course (N = 5), the arterioles and venules presented transient increases of 6.5 and 5.2% in caliber, respectively (Fig. 5 g); however, a minimal change of 2.38% was observed in the decorrelation value (mean ± std: baseline 0.42 ± 0.26 vs. FLS 0.43 ± 0.35). Moreover, considerably distinct to the arterioles and venules, the SVP capillaries, ICP, and DCP showed a maximal increase of 9.7, 7.4, and 3.3% in decorrelation during FLS, respectively (Fig. 5 h), which may indicate the differences in nutrient supply in different vessels.
Fig. 5. FLS-evoked hemodynamic responses of mouse retina in vivo. Representative en face angiograms of SVP, ICP and DCP slabs in baseline (a-c) and FLS (d-f) phases. Decorrelation value were encoded with color. Insets (I-VIII) are the enlarged views of the local regions that marked by red rectangles in (a-f). The red dashed curves in (I and II) indicate the vessel profile under stimulus. The red triangles in (III-VIII) indicate the apparent vascular changes. Time courses of vessel caliber (g) and decorrelation signal (h). Arterioles and venules were differentiated empirically based on their morphological differences in the en face OCTA image [24], and the vessel caliber was averaged over all arterioles and venules separately. FLS Off in (g and h) means the dynamic changes of the vessels and blood flow when FLS is turned off. Each dynamic OCTA sequence (5 baseline and 5 FLS volumes) used the threshold determined by the first baseline volume to eliminate the influence of different thresholds. SVP, superficial vessel plexus in the ganglion cell layer; ICP, intermediate capillary plexus in the inner plexiform layer; DCP, deep capillary plexus in the outer plexiform layer.
4. Discussion and conclusion
With the advancement of the OCTA imaging speed, the current 133 kHz line-processing rate is expected to be challenged again by faster OCTA systems. As shown in Table 1, ∼ 69% of the processing time was consumed by the tasks of alignment and segmentation, which can be disabled for the initial real-time feedback of the vascular network visibility during data acquisition. Accordingly, the ID-OCTA processing speed was further boosted to a line rate of ∼ 430 kHz to satisfy the fast 400 kHz preclinical system [35]. In addition, the maximal line-processing rate was also affected by the setting of the batch size in the GPU pipeline, which should be optimized to maintain a balance between data acquisition, transmission, and processing.
Noise artifacts severely degraded the angiogram quality, which is commonly suppressed by thresholding [6,36]. SNR-adaptive thresholding was achieved under the guidance of a derived asymptotic linear ID relationship [9]. However, because of the limited kernel size used in decorrelation estimation in practice, the ID relation was not convergent, leading to a considerable intermediate region ${R_i}\; $ with a mixed distribution of static and dynamic voxels. Rather than completely removing region ${R_i}$, the ambiguous voxels in the intermediate region could be further differentiated with a shape mask [19]; however, it was computationally intensive and difficult to achieve with GPU parallel computation. Thus, as a compromise, the ID threshold was optimized in this study by maximizing the angiogram quality in the retina, that is, a trade-off between reserving flow voxels and removing background voxels in region ${R_i}$. Consequently, the ID threshold auto-optimization was readily accomplished within 0.12 s without obvious degradation of image quality owing to the real-time data processing in GPU ID-OCTA.
Several metrics have been developed for evaluating the quality of OCTA image [31,33,37]. Considering the computational complexity and cost, the SNR, CNR, and CON metrics were used for the final validation of the improved performance of the optimized threshold, and a low-cost quality index $QI$ ($\gamma $) was used as the threshold feedback by combining the vessel number $VD$ and image contrast ${C_{RMS}}$. For the quality index, the curves of $VD$ ($\gamma $) and ${C_{RMS}}$ ($\gamma $) were normalized to reduce the influence of the raw signal intensity and vessel density. In addition, the search precision of the optimal ${\gamma _T}$ can be further improved by increasing the number of iterations M and by turning the weights of $VD$ and ${C_{RMS}}$ in $QI$.
The FLS-evoked hemodynamic changes can be attributed to the increased metabolic demands in response to the evoked neural activities [38–40], and these changes were consistent with the reported literature. The measured ∼5.2–6.5% vasodilation of arterioles and venules in the mouse retina was consistent with the reported ∼4.9–6.8% in the human retina [41], and the observed ∼3.3–9.7% elevation of decorrelation value in capillary plexuses was close to the reported ∼5.4–6.8% increase during light adaptation of the human retina [38]. Moreover, the minimal change of 2.38% in the arterioles and venules decorrelation value (mean ± std: baseline 0.42 ± 0.26 vs. FLS 0.43 ± 0.35) was probably because of the saturation limit (mean ± std: 0.56 ± 0.20 for a kernel size of 5) in the decorrelation estimation [10,42]. Furthermore, the dynamic decorrelation change was suggested to be an index of blood flux, which is a combination of the speed and density of RBCs [10].
The ability to visualize FLS-evoked hemodynamics is of great importance for investigating neurovascular coupling and evaluating inner retinal neurodegeneration, which is a preclinical sign of diabetic retinopathy [43–45]. Although the dynamic vessel analyzer could evaluate endothelial dysfunction by analyzing the flicker-evoked dilation of retinal arterioles and venules, it required considerable imaging time of 350 s per eye and had no capillary resolution [41,46]. Whereas, fluorescence-based confocal microscopy could detect hemodynamic responses of blood flow signals down to the capillary level [45], but the requirement of exogenous contrast agents and the limited imaging field (micron scale) hindered its clinical application in practice. In contrast, the proposed GPU-accelerated dynamic ID-OCTA enabled fast and label-free 4D (3D space and time) visualization of FLS-evoked hemodynamics and quantification of both vessel caliber and decorrelation signals within a single trail (∼60 s).
In summary, a GPU-based data-processing pipeline was developed for dynamic ID-OCTA imaging. The pipeline had a measured line-process rate of 133 kHz for displaying the OCT and OCTA cross-sections in real time. Further, real-time processing enabled automatic optimization of angiogram quality, which improved the SNR, CNR, and vessel connectivity by 14.37, 14.08, 9.76%, respectively. Furthermore, motion-contrast 4D angiographic imaging of FLS-evoked hemodynamics was achieved within a single trail in the mouse retina. Accordingly, the FLS evoked an apparent dilation of the arterioles and venules in the SVP, and an elevation of the decorrelation value in all retinal plexuses. Therefore, GPU ID-OCTA enables real-time and high-quality angiographic imaging and is particularly suitable for hemodynamic studies.
Funding
National Natural Science Foundation of China (11974310, 31927801, 62035011, 62075189); Zhejiang Provincial Natural Science Foundation of China (LR19F050002); Zhejiang Lab (2018EB0ZX01); National Key Research and Development Program of China (2017YFA0700501); the MOE Frontier Science Center for Brain Science & Brain-Machine Integration, Zhejiang University.
Disclosures
The authors declare no conflicts of interest.
Data Availability
Data underlying the results presented in this paper are not publicly available at this time but may be obtained from the authors upon reasonable request. Correspondence should be addressed to the corresponding author.
References
1. C. L. Chen and R. K. Wang, “Optical coherence tomography based angiography [Invited],” Biomed. Opt. Express 8(2), 1056–1082 (2017). [CrossRef]
2. J. Fingler, D. Schwartz, C. Yang, and S. E. Fraser, “Mobility and transverse flow visualization using phase variance contrast with spectral domain optical coherence tomography,” Opt. Express 15(20), 12636–12653 (2007). [CrossRef]
3. A. Mariampillai, B. A. Standish, E. H. Moriyama, M. Khurana, N. R. Munce, M. K. K. Leung, J. Jiang, A. Cable, B. C. Wilson, I. A. Vitkin, and V. X. D. Yang, “Speckle variance detection of microvasculature using swept-source optical coherence tomography,” Opt. Lett. 33(13), 1530–1532 (2008). [CrossRef]
4. B. J. Vakoc, R. M. Lanning, J. A. Tyrrell, T. P. Padera, L. A. Bartlett, T. Stylianopoulos, L. L. Munn, G. J. Tearney, D. Fukumura, R. K. Jain, and B. E. Bouma, “Three-dimensional microscopy of the tumor microenvironment in vivo using optical frequency domain imaging,” Nat. Med. 15(10), 1219–1223 (2009). [CrossRef]
5. A. Zhang, Q. Zhang, C. L. Chen, and R. K. Wang, “Methods and algorithms for optical coherence tomography-based angiography: a review and comparison,” J. Biomed. Opt. 20(10), 100901 (2015). [CrossRef]
6. J. Enfield, E. Jonathan, and M. Leahy, “In vivo imaging of the microcirculation of the volar forearm using correlation mapping optical coherence tomography (cmOCT),” Biomed. Opt. Express 2(5), 1184–1193 (2011). [CrossRef]
7. L. Guo, P. Li, C. Pan, R. Liao, Y. Cheng, W. Hu, Z. Chen, Z. Ding, and P. Li, “Improved motion contrast and processing efficiency in OCT angiography using complex-correlation algorithm,” J. Opt. 18(2), 025301 (2016). [CrossRef]
8. A. H. Kashani, C.-L. Chen, J. K. Gahm, F. Zheng, G. M. Richter, P. J. Rosenfeld, Y. Shi, and R. K. Wang, “Optical coherence tomography angiography: A comprehensive review of current methods and clinical applications,” Prog. Retinal Eye Res. 60, 66–100 (2017). [CrossRef]
9. L. Huang, Y. Fu, R. Chen, S. Yang, H. Qiu, X. Wu, S. Zhao, Y. Gu, and P. Li, “SNR-Adaptive OCT Angiography Enabled by Statistical Characterization of Intensity and Decorrelation With Multi-Variate Time Series Model,” IEEE Trans. Med. Imaging 38(11), 2695–2704 (2019). [CrossRef]
10. R. Chen, L. Yao, K. Liu, T. Cao, H. Li, and P. Li, “Improvement of Decorrelation-Based OCT Angiography by an Adaptive Spatial-Temporal Kernel in Monitoring Stimulus-Evoked Hemodynamic Responses,” IEEE Trans. Med. Imaging 39(12), 4286–4296 (2020). [CrossRef]
11. V. J. Srinivasan, H. Radhakrishnan, E. H. Lo, E. T. Mandeville, J. Y. Jiang, S. Barry, and A. E. Cable, “OCT methods for capillary velocimetry,” Biomed. Opt. Express 3(3), 612–629 (2012). [CrossRef]
12. J. Yang, J. Su, J. Wang, S. Men, Y. Jia, D. Huang, and G. Liu, “Hematocrit dependence of flow signal in optical coherence tomography angiography,” Biomed. Opt. Express 8(2), 776–789 (2017). [CrossRef]
13. P. Li, S. Yang, Z. Ding, and P. Li, “Photon ensemble correlation spectroscopy enables decorrelation-rate-limited ultrafast measurement of diffusive particle dynamics,” Opt. Lett. 42(21), 4525–4528 (2017). [CrossRef]
14. G. Liu, W. Jia, V. Sun, B. Choi, and Z. Chen, “High-resolution imaging of microvasculature in human skin in-vivo with optical coherence tomography,” Opt. Express 20(7), 7694–7705 (2012). [CrossRef]
15. J. Tokayer, Y. Jia, A. H. Dhalla, and D. Huang, “Blood flow velocity quantification using split-spectrum amplitude-decorrelation angiography with optical coherence tomography,” Biomed. Opt. Express 4(10), 1909–1924 (2013). [CrossRef]
16. J. P. Su, R. Chandwani, S. S. Gao, A. D. Pechauer, M. Zhang, J. Wang, Y. Jia, D. Huang, and G. Liu, “Calibration of optical coherence tomography angiography with a microfluidic chip,” J. Biomed. Opt. 21, 1 (2016). [CrossRef]
17. X. Wei, A. Camino, S. Pi, T. T. Hormel, W. Cepurna, D. Huang, J. C. Morrison, and Y. Jia, “Real-time cross-sectional and en face OCT angiography guiding high-quality scan acquisition,” Opt. Lett. 44(6), 1431–1434 (2019). [CrossRef]
18. J. Xu, S. Han, C. Balaratnasingam, Z. Mammo, K. S. K. Wong, S. Lee, M. Cua, M. Young, A. Kirker, D. Albiani, F. Forooghian, P. Mackenzie, A. Merkur, D.-Y. Yu, and M. V. Sarunic, “Retinal angiography with real-time speckle variance optical coherence tomography,” Br. J. Ophthalmol. 99(10), 1315–1319 (2015). [CrossRef]
19. H. K. Li, K. Y. Liu, T. T. Cao, L. Yao, Z. Y. Zhang, X. F. Deng, C. X. Du, and P. Li, “High performance OCTA enabled by combining features of shape, intensity, and complex decorrelation,” Opt. Lett. 46(2), 368–371 (2021). [CrossRef]
20. Y. Jian, K. Wong, and M. V. Sarunic, “Graphics processing unit accelerated optical coherence tomography processing at megahertz axial scan rate and high resolution video rate volumetric rendering,” J. Biomed. Opt. 18(02), 1 (2013). [CrossRef]
21. S. Borkovkina, A. Camino, W. Janpongsri, M. V. Sarunic, and Y. Jian, “Real-time retinal layer segmentation of OCT volumes with GPU accelerated inferencing using a compressed, low-latency neural network,” Biomed. Opt. Express 11(7), 3968 (2020). [CrossRef]
22. J. Xu, K. Wong, Y. Jian, and M. Sarunic, “Real-time acquisition and display of flow contrast using speckle variance optical coherence tomography in a graphics processing unit,” J. Biomed. Opt. 19, 026001 (2014). [CrossRef]
23. C. L. Chen, W. S. Shi, and V. X. D. Yang, “Real-time en-face Gabor optical coherence tomographic angiography on human skin using CUDA GPU,” Biomed. Opt. Express 11(5), 2794–2805 (2020). [CrossRef]
24. K. Y. Liu, T. P. Zhu, L. Yao, Z. Y. Zhang, H. K. Li, J. Ye, and P. Li, “Noninvasive OCT angiography-based blood attenuation measurements correlate with blood glucose level in the mouse retina,” Biomed. Opt. Express 12(8), 4680–4688 (2021). [CrossRef]
25. ANSI, American National Standard for Safe Use of Lasers, ANSI Z136.1–2007 (Laser Institute of America, Orlando, 2007), pp. 1–249.
26. X. Wei, T. T. Hormel, Y. K. Guo, T. S. Hwang, and Y. L. Jia, “High-resolution wide-field OCT angiography with a self-navigation method to correct microsaccades and blinks,” Biomed. Opt. Express 11(6), 3234–3245 (2020). [CrossRef]
27. T. Fabritius, S. Makita, M. Miura, R. Myllyla, and Y. Yasuno, “Automated segmentation of the macula by optical coherence tomography,” Opt. Express 17(18), 15659–15669 (2009). [CrossRef]
28. E. Peli, “Contrast in Complex Images,” J. Opt. Soc. Am. A 7(10), 2032–2040 (1990). [CrossRef]
29. T. T. Hormel, J. Wang, S. T. Bailey, T. S. Hwang, D. Huang, and Y. L. Jia, “Maximum value projection produces better en face OCT angiograms than mean value projection,” Biomed. Opt. Express 9(12), 6412–6424 (2018). [CrossRef]
30. M. S. Bajwa, A. P. Agarwal, and S. Manchanda, “Ternary Search Algorithm: Improvement of Binary Search,” 2015 2nd International Conference on Computing for Sustainable Global Development (Indiacom), 1723–1725 (2015).
31. Y. L. Jia, O. Tan, J. Tokayer, B. Potsaid, Y. M. Wang, J. J. Liu, M. F. Kraus, H. Subhash, J. G. Fujimoto, J. Hornegger, and D. Huang, “Split-spectrum amplitude-decorrelation angiography with optical coherence tomography,” Opt. Express 20(4), 4710–4725 (2012). [CrossRef]
32. N. Otsu, “A Threshold Selection Method from Gray-Level Histograms,” IEEE Trans. Syst., Man, Cybern. 9(1), 62–66 (1979). [CrossRef]
33. J. Xu, S. Song, Y. Li, and R. K. Wang, “Complex-based OCT angiography algorithm recovers microvascular information better than amplitude- or phase-based algorithms in phase-stable systems,” Phys. Med. Biol. 63(1), 015023 (2017). [CrossRef]
34. Q. Zhu, X. Xing, M. Zhu, H. Xiao, L. Ma, L. Chen, J. Liang, Y. Yuan, and E. Song, “A New Approach for the Segmentation of Three Distinct Retinal Capillary Plexuses Using Optical Coherence Tomography Angiography,” Trans. Vis. Sci. Tech. 8(3), 57 (2019). [CrossRef]
35. S. Ni, X. Wei, R. Ng, S. Ostmo, M. F. Chiang, D. Huang, Y. Jia, J. P. Campbell, and Y. Jian, “High-speed and widefield handheld swept-source OCT angiography with a VCSEL light source,” Biomed. Opt. Express 12(6), 3553–3570 (2021). [CrossRef]
36. W. J. Choi, R. Reif, S. Yousefi, and R. K. K. Wang, “Improved microcirculation imaging of human skin in vivo using optical microangiography with a correlation mapping mask,” J. Biomed. Opt. 19(3), 036010 (2014). [CrossRef]
37. Z. Chu, J. Lin, C. Gao, C. Xin, Q. Zhang, C.-L. Chen, L. Roisman, G. Gregori, P. J. Rosenfeld, and R. K. Wang, “Quantitative assessment of the retinal microvasculature using optical coherence tomography angiography,” J. Biomed. Opt. 21(6), 066008 (2016). [CrossRef]
38. P. L. Nesper, H. E. Lee, A. E. Fayed, G. W. Schwartz, F. Yu, and A. A. Fawzi, “Hemodynamic response of the three macular capillary plexuses in dark adaptation and flicker stimulation using optical coherence tomography angiography,” Invest. Ophthalmol. Visual Sci. 60(2), 694–703 (2019). [CrossRef]
39. M. T. Wong-Riley, “Energy metabolism of the visual system,” Eye Brain 2, 99–116 (2010). [CrossRef]
40. Y. S. Zhang, I. Mucollari, C. C. Kwan, G. Dingillo, J. Amar, G. W. Schwartz, and A. A. Fawzi, “Reversed neurovascular coupling on optical coherence tomography angiography is the earliest detectable abnormality before clinical diabetic retinopathy,” J. Clin. Med. 9(11), 3523 (2020). [CrossRef]
41. M. Lim, M. B. Sasongko, M. K. Ikram, E. Lamoureux, J. J. Wang, T. Y. Wong, and C. Y. Cheung, “Systemic associations of dynamic retinal vessel analysis: a review of current literature,” Microcirculation 20(3), 257–268 (2013). [CrossRef]
42. Y. Nakamichi, K. S. Chiu, and C. W. Sun, “Signal properties of split-spectrum amplitude decorrelation angiography for quantitative optical coherence tomography-based,” Biomed. Opt. Express 12(10), 5955–5968 (2021). [CrossRef]
43. C. N. Hall, C. Reynell, B. Gesslein, N. B. Hamilton, A. Mishra, B. A. Sutherland, F. M. O’Farrell, A. M. Buchan, M. Lauritzen, and D. Attwell, “Capillary pericytes regulate cerebral blood flow in health and disease,” Nature 508(7494), 55–60 (2014). [CrossRef]
44. A. Duan, P. A. Bedggood, B. V. Bui, and A. B. Metha, “Evidence of flicker-induced functional hyperaemia in the smallest vessels of the human retinal blood supply,” PLoS ONE 11(9), e0162621 (2016). [CrossRef]
45. T. E. Kornfield and E. A. Newman, “Regulation of blood flow in the retinal trilaminar vascular network,” J. Neurosci. 34(34), 11504–11513 (2014). [CrossRef]
46. C. Y. Cheung, M. K. Ikram, R. Klein, and T. Y. Wong, “The clinical implications of recent studies on the structure and function of the retinal microvasculature in diabetes,” Diabetologia 58(5), 871–885 (2015). [CrossRef]
