1. Introduction

Glaucoma is the second leading cause of blindness in patients with eye diseases globally. With the progression of glaucoma, the peripheral vision and central vision may decay in sequence, causing permanent blindness if not treated timely [1]. A number of biomarkers have been studied as possible causal factors in glaucoma. As a main characteristic of glaucoma, elevated intraocular pressure (IOP) has a detrimental effect in the progression of this disease, but the relation between IOP and optic nerve damage varies among patients. Some patients will show ophthalmologic evidence of glaucoma with normal IOPs [2]. Multiple biomarkers of glaucoma have been investigated as possible diagnostic features of glaucoma [3,4], including alterations in biomechanical properties of the optic nerve head (ONH) [5], prolonged eye wall stress, local vascular abnormalities, and decreased ocular perfusion pressure. Retinal ganglion cells (RGCs) are the first damaged cells in glaucoma and play an important role in the management of visual information. When light enters the eye, the RGC axons will transmit the signals to the brain. Studies have also revealed that retaining the RGCs and axons in an early stage of glaucoma is quite necessary for sight [6,7]. The RNFL consists of RGC axons, representing the innermost layer of the fundus. Hence, to find evidence about abnormal changes of early glaucoma, it is indispensable to detect the progression of RGC axon loss within the RNFL in glaucoma. However, to image the RGCs, it requires ultra-high resolution (∼2 µm) imaging system, which is usually costly and complicated to build. Alternatively, growing evidence has suggested that RNFL loss precedes visual field loss in glaucoma from macroscopic perspective. [8]. Therefore, to enhance the abilities of detecting damage in early glaucoma, evaluations of the RNFL become significant. It has been reported that a RNFL defect (RNFLD) appears earlier than loss of the optic disk and visual field [9]. A RNFLD has been demonstrated in ∼60% of the patients 6 years before loss of the visual field. The location of a RNFLD was highly correlated with the location of subsequent visual field loss. In the early stage of glaucoma, the morphology of the RNFL has not changed significantly, but the cells of RNFL might have been damaged, which could be evaluated through analyzing the elasticity of RNFL. The change in the elasticity of the RNFL may affect the waveform change of the pulsation of the RNFL, such as the amplitude. Therefore, extracting the pulsations of RNFL, analyzing the waveform changes of RNFL, and evaluating RNFLD are very helpful for the early diagnosis and intervention of glaucoma.

Currently, as the thickness of RNFL decreases gradually in glaucoma patients, changes in RNFL thickness measured using traditional methods remain prevalent. However, changes in RNFL thickness are too subtle to be detected in early glaucoma. Studies have indicated that the available routine examination modalities cannot detect functional damage at an early stage of glaucoma. The serial RNFL examination was more sensitive than color disc evaluation in the detection of progressive damage in early glaucoma. Here, the RGCs decay induced RNFL dynamic motion abnormality is assumed to be more sensitive than the RNFL morphological change in early glaucoma. Changes in RNFL dynamic data may occur earlier than the thickness of RNFL. As one of the basic characteristics, visual impulses originating from the rods and cones are integrated by the RNFL. Travelling through the RGC, the visual impulses are transmitted from the rods and cones to the RNFL. The pulsatile movements of the ocular fundus might contribute to axonal damage and loss of the RNFL in glaucoma. Measurement of the RNFL impulses can contribute greatly to the early diagnosis of glaucoma.

Studies in evaluating the pulsatile movement of the ONH and the peripapillary retina in normal subjects and glaucoma patients have revealed that axonal damage and loss of the RNFL in glaucoma might be attributed to the pulsation of the ocular fundus. The pulsation might cause damage to the optic disc and RNFL in the progression of early glaucoma [10]. Studies in measuring the RNFL thickness and spontaneous retinal venous pulsations in glaucoma patients and normal subjects [11] have also determined that the amplitude of retinal venous pulsations is reduced in glaucoma with growing RNFL loss. The latest studies pointed out that biomechanical properties and ocular hemodynamics may be the other two key characteristics of the pathophysiology of glaucoma [12].

Here, we postulate the progression of retinal impulses might be a dynamic marker for glaucoma severity. Detection of pulse-induced movement of the RNFL might provide a breakthrough for early diagnosis of glaucoma. Because functional measurement tools have been lacking, no currently available technology is capable of measuring RNFL movement, probably because it is too small (typically a few micrometers). It requires that the current available technology is able to measure RNFL movement, which is within a few micrometers.

Evaluation methods with non-invasive approaches for measuring dynamic characteristics in vivo are more useful for a complete understanding of the mechanisms of nerve cell degeneration with this disease. Optical coherence tomography (OCT) has become a powerful imaging modality in biomedical and clinical studies [13,14] during decades of development. It has produced great achievement in ophthalmology by incorporating retinal and choroidal structural and functional imaging. Since the axial resolution is decoupled with lateral resolutions, OCT provides more superior axial sectional slabs. A recent study has shown that structural-OCT may be used to evaluate the pulsatile movement of ONH by analyzing the alternation of distances between the cornea and ONH [10]. However, the axial resolution of the reported system was relatively low (7 µm), while the measured maximal ONH movement was only ∼10 µm for normal subjects and ∼14 µm for glaucoma subjects. It is difficult for the system to distinguish axial motion specific to the RNFL from the axial motion of the entire globe. The custom-made phase-sensitive OCT apparatus is used to measure micrometer scale movement of fundus tissue and isolate the RNFL component using time-lapse B-scans. Aiming to collect the phase information from adjacent A-scan signals, phase sensitive optical coherence tomography (PhS-OCT) was first reported to measure the Doppler effect of tissues [15]. However, the evaluation of phases between adjacent A-scans is not amenable to the measurement of the tissue dynamic motion within micrometer scale. To overcome the difficulty, PhS-OCT was further developed by evaluating phases between adjacent B-scans. The enlarged increment of the time interval between the B-scans results in sensitivity enhancement on a nanometer scale [16]. An et al. [17] also used an ultra-high-speed SD-OCT system to measure the axial motion with continuous oscillations of the ONH, demonstrated that the ONH is closely associated with pulsative blood flow simultaneously measured. Longitudinal movements of the retina and cornea in living rats were successfully measured by the frequency domain optical coherence tomography (FD-OCT) system [12], results showed that these two movements are almost with the same synchrony and direction both in the systolic period and diastolic period, it can be used to assess the biomechanical properties of the eye and have important implications for early diagnosis, pathophysiological studies of glaucoma and other eye diseases.

In 2010, Yu et al. [18] proposed a histogram-based method for body motion correction in variance imaging, demonstrated that histogram analysis works well for batch motion artifact correction. In 2013, an algorithm [19] based on two-dimensional cross-correlation is developed to correct the lateral and axial displacement between B-scans generated by the motion of the eyes. The time interval between the adjacent frames is used to evaluate fundus pulsation, the retina phase difference is used to compensate the axial motion. This method does not compensate for lateral or rotational eye motions. In 2018, an algorithm [20–21] based on extracting phase information from the acquired time-resolved OCT datasets is proposed, which aligned the adjacent B-scans precisely and automatically to minimize the unexpected error caused by pixel mismatch.

However, there are some problems with the measurement of the pulsatile RNFL movements using OCT. Most of prior art focus on the extraction of anesthetized animals, or on the extraction of trabecular meshwork in vivo, it is still a challenge to extract of RNFL pulsation parameter in vivo. Inevitable movements usually result in a bulk tissue dynamic imaging artifact. To overcome the difficulty, a novel algorithm is proposed by using phase measurement from the tissue near the RNFL as a reference to compensate for the bulk-tissue motion artifact. Afterwards, the phase alterations due to RNFL pulsatile movements are extracted.

Here, we measured RNFL tissue movements induced by the cardiac pulsation. In this process, the ocular pulse waveform as an indicator of tissue compliance was studied. Using a non-invasive SD-OCT technique with high resolution and imaging speed, overcome interference of body movement and keep the B-scan correlation and stability while detecting ocular hemodynamic and biomechanical properties. The cross-sectional images involving the pulsatile movement of the RNFL were measured. The cardiac pulse waveform and the RNFL pulse waveform are compared and analyzed in time and frequency domain to validate the feasibility and accuracy of the in vivo extraction scheme. The spectral analysis of the fundus pulsation is a dynamic indicator for ocular biomechanical properties. Hence, pulsatile movement of the RNFL could be a biomarker as one of the determinant parameters of the degeneration of RGCs and subsequent RNFLD in glaucoma.

2. Methods

2.1 Subjects

The healthy subjects with no history of ocular disease were recruited, and this experiment had collected the left and right eye data. None of the subjects were using ocular medications and none of them had a history of ocular disease, laser procedures or intraocular surgery. All volunteers signed an informed consent form before participating in the experiment.

2.2 SD-OCT system setup

The SD-OCT performed here was similar to that previously reported [22]. The system schematic is delineated in Fig. 1. The system is based on low-coherence interferometry using a broadband infrared laser diode at a center wavelength of 840 nm with bandwidth of ±49 nm, giving an axial resolution of 6.3µm in air. The laser beam is split into two pathways by a 50/50 coupler fiber, one within a reference arm and one directed towards the eye within a sample arm. In the sample arm, a lens group is assembled before the eye, the beam is focused on the fundus through the lens group. The OCT image is generated using light back-reflected from these two pathways. These two reflections form an interference pattern which is subsequently analyzed with spectrometer.

 

Fig. 1. The schematic of SD-OCT system.

Download Full Size | PDF

A high-speed spectrometer detection unit was employed in the system, which could provide a 120 kHz A-lines rate for 1024 pixels, the scanning speed can be up to 20-400 B-scans per second. The sensitivity was calculated to be 80 dB for a single image at the speed of 120 kHz. The collimated beam had a diameter of 3 mm, giving a lateral resolution of 15 µm.

2.3 Data acquisition

2.3.1 OCT image

The primary requirement of the OCT scanning region should include both the RNFL and peripapillary retina to successfully extract the RNFL motion within a micro-scale. The imaging rate is 400 frames per second (fps). In each dataset, 2000 B-scans were acquired within ∼5 s to cover 5 heart pulse cycles, according to the human heart rate fundamental frequency being ∼1 Hz.

The schematic diagrams of the location and range of SD-OCT imaging of the retina are shown in Fig. 2. Before imaging RNFL pulsatile movements, a fundus image was captured with LSO (Line Scanning Ophthalmoscope, LSO) scan of the region expanding from the fovea. The result is shown in Fig. 2(a). Pulsatile RNFL movements and blood flow located in the blood vessels of the ONH were measured. Using a protocol of repeated B-scan time-lapsed mode, 300 A-lines of 1024 pixels were captured to form one B-scan (covering ∼3 mm in length), the position of B-scan is indicated by the red line in Fig. 2(a), which can be adjusted according to the fundus retina map displayed in the LSO map. In practice, we first find the central artery near the optic nerve head in the LSO map, and adjust the OCT acquisition frame to the red line shown in Fig. 2(a), the collected is the B-scan image of the corresponding position, and each position repeats 2000 times B-scan. Figure 2(b) is a B-scan image of the fundus retina corresponding to the acquisition position shown in the LSO image. The red box is the scanned area required actually in the experiment, the green dashed box is the location of the algorithm processing and the blood vessel near the scanned area is marked by the yellow box.

 

Fig. 2. Schematic diagram of retinal scanning range. (a) LSO diagram, the red line is the OCT scan position; (b) B-scan diagram, the red box is the scan position actually required, the green dashed box is the location of the algorithm processing, and the yellow box is the blood vessel near the required scan position.

Download Full Size | PDF

Before conducting the in vivo experiments, informed consent was obtained from the subject, and the experimental procedures adhered to the tenets of the Declaration of Helsinki. When the SD-OCT performing real-time line scanning of the retina, the human eye was aligned in the sample arm along the optical axis with the temples and chin pressed against the headrest to minimize head movements, and a fixation target was used to reduce lateral eye motion. The probe power of the OCT is less than or equal to 600 µW, powers applied to the retina were below the maximum permissible exposure according the American national standards institute (ANSI) and international organization for standardization (ISO) standards. Note that by minimizing the light level in the room, the subject had a pupil of 4.5 mm during the experiment.

2.3.2 Cardiac pulse

An infrared pulse sensor device of HKG-07C based on LabVIEW program control measures the heart rate at the same time when collecting OCT signals.

2.4 Algorithm

2.4.1 Image segmentation algorithm

An image segmentation algorithm based on the shortest path and convolutional neural network is proposed to realize the segmentation between the outer membranes of the inner bound membrane. It includes two steps. First, the Dijkstra algorithm proposed by Chiu [23] et al. is used to segment images. The segmented images were used as standard datasets to train the convolutional neural networks. Second, by training on a standard dataset, a convolutional neural network model for RNFL segmentation has been segmented for a large number of retinal OCT images of different sizes at different sites. The final segmentation result is also estimated by comparison with manual segmentation. The purpose of the segmentation is to remove the interfering tissues around for subsequent phase extraction. For better segmentation, we use two algorithms of segmentation method. Dijkstra's algorithm used to get the standard segmented image to train the neural network. Convolutional neural network used to segment a large number of data.

The shortest path algorithm is a method of calculating the shortest path from a node to all other nodes. Since the characteristics of each layer of the OCT retinal B-scan image are obvious, the shortest path algorithm can be used for the layering of the OCT retinal image. The shortest path hierarchical algorithm theory used in this paper is a fully automatic visual image hierarchical structure segmentation method using graph theory and dynamic programming [23]. The workflow is shown in Fig. 3. The first cross-sectional image in Fig. 3 is a B-scan of the retina acquired by OCT; the second cross-sectional image in Fig. 3 is obtained from the first cross-sectional image. A contour extraction-based method was used to flatten the retina. First, the profile of the retinal image was detected by threshold binarization, then, the retinal radian was obtained from the profile, and finally, the retinal image was flattened by adding zero at the beginning or end of the column vector of the image matrix to improve the accuracy of finding the shortest path. The flattened B-scan image of the retina was inputted, and then the gradient weight of vertical direction was calculated. The pixel at each point in the input B-scan image is used as the node in the shortest path algorithm. Since the retinal image collected in this experiment is mainly a horizontal structure, the weight value in the shortest path algorithm is selected to calculate by the gradient of pixel intensity change in the vertical direction. Then, the layer endpoints are initialized, one column node on each side of the image is added to restrict the search area and set the invalid node. The purpose of restricting the search area is to avoid repeated searches in the divided area, and remove the weight of the edge in the invalid node. Finally, the minimum weighted path is found by summing the weights around the starting node to get the minimum. The shortest path between each pixel obtained by weight is the required segmentation line. The third cross-sectional view in Fig. 3 shows the final overall retinal segmentation results.

 

Fig. 3. Work flow of the shortest path algorithm.

Download Full Size | PDF

The collected retinal B-scan images were segmented by the shortest path algorithm to be used as the standard dataset to train the convolutional neural network. 2000 retinal B-scan images with segmentation and calibration in the standard dataset are divided into three parts - training set, verification set and test set, with the proportion of 8:1:1 to input into the convolutional neural network to training the model.

The specific structure of the convolutional neural network used in our study is shown in Fig. 4. It is a full convolutional neural network, including the input layer, Conv1∼Conv5, a total of 15 convolution layers. The size of the convolution kernel is 3×3, and rectified linear unit (ReLU) is used as the excitation function. The number of channels in the convolutional feature layer of Conv1∼Conv5 is 32, 64, 128, 256, and 3, respectively.

 

Fig. 4. (a) Convolutional neural network structure; (b) Convolutional neural network segmentation result; (c) Manual segmentation result. The boundary marked by the blue line is the inner boundary membrane, the yellow line is the outer membrane.

Download Full Size | PDF

The hardware device for the image segmentation algorithm is a computer with two Intel E5-2678V3 CPU, 64G memory and four NVIDIA GTX1080 8G graphics cards. The algorithm code is implemented by Python 3.6 in the Linux operating system, The neural network model is based on open source Tensorflow 1.14 deep learning framework using computer parallel GPU physical acceleration during training. Model initialization weights were performed by Xavier initialization, adaptive moment estimation (Adam) estimation was used as the optimizer during training with initial learning rate set to 0.001, maximum iterations set to 400000 and training batch set to 4.

To test the trained neural network, a total of 2000 frames were collected from another volunteer and re-inputted into the full convolutional neural network, the output retinal segmentation results, as shown in Fig. 4(b), are compared with Manual segmentation result (Fig. 4(c)). In the figures, the boundary marked by the blue line is the inner boundary membrane, and the position marked by the yellow line is the outer membrane. Two hundred of the two thousand frames are selected randomly, segmented manually, and then the manual segmentation results were used as standard to calculate the index of accuracy, precision, recall and MUE (Mean unsigned error). After calculation, the convolutional neural network segmentation has an accuracy of 99.26%, a precision of 97.51%, and a recall of 97.84%. The MUE value of the inner boundary membrane segmentation line was calculated as 1.65 pixels and the outer membrane segmentation line as 5.67 pixels. The results show that the images can be segmented effectively by the trained neural network. The purpose of the segmentation is to remove the interfering tissues around for subsequent phase extraction.

2.4.2 Phase extraction

Since histogram analysis is effective in batch motion artifact correction, we used the histogram algorithm proposed by Yu et al. [18]. A phase extraction algorithm for RNFL motion is proposed by combining the histogram phase extraction algorithm with two body motion phase compensation algorithms. The workflow of the phase extraction algorithms is shown in Fig. 5. The first cross-sectional image in Fig. 5 is a B-scan of the retina acquired by OCT. The second cross-sectional image in Fig. 5 is the phase difference plot obtained after calculating the phase difference between the first B-scan plot and the adjacent next B-scan plot. Using a protocol of repeated B-scan time-lapsed mode, 300 A-lines of 1024 pixels were captured to form one B-scan (covering ∼3 mm in length). The line number in x-axis refers to the sequence number of the A-lines in the OCT B-scans.

 

Fig. 5. Flow chart of phase extraction algorithm.

Download Full Size | PDF

First, the segmentation line between the outer membrane and the inner membrane was found by the aforementioned image segmentation algorithm, and the original spectral data of the part to be processed were separated by using the segmentation line. That means, the input of the algorithm is the original OCT spectral signal with phase information between the outer membrane and the inner membrane instead of the image. After the data is input, the total phase difference between two adjacent frames is calculated, including body motion and tissue micro motion. The histogram of each A-line in the phase difference map is calculated to generate the histogram curve, and then judge whether the generated histogram curve includes phase wrapping. If so, unwrap the phase wrapping. The distal end of the curve without phase wrapping or unwrapped is fitted by the first-order polynomial. Then the fitted curve is subtracted by the total phase difference curve to get the corrected phase difference curve.

Taking the adjacent 67^th and 68^th frames images as example, Fig. 6 represents the phase difference curve before and after fitting the subtraction. The line number in x-axis refers to the sequence number of the A-lines in the OCT B-scans. The blue line in Fig. 6 represents the total phase difference curve between the 67^th and 68^th frames. The part far away from the large blood vessels is regarded as the stationary end, in practice, the first 100 A-lines is regarded as the stationary end, that is, the first 100 points in the total phase difference curve in Fig. 6 are fitted with first-order polynomials, and the fitting curve is the red line in Fig. 6. The yellow line in Fig. 6 shows the final phase difference curve after subtracting polynomial fitting from the total phase difference curve. The last 200 points in the final phase difference curve are the phases generated at the moving end of the retinal nerve fiber layer after removing the phase generated by body motion. The phase difference can be used to characterize the displacement within 2.5ms, which is used to analyze RNFL motion in the next step.

 

Fig. 6. Phase difference curves before and after fitting subtraction.

Download Full Size | PDF

3. Results

According to the scan range of the retina shown in Fig. 2, there are a total of 2000 B-scan images in a dataset collected by SD-OCT, 1999 phase difference curves are yielded after phase extraction. Figure 7 shows a graph of the phase difference over time, which plots the 1999 phase difference curves together. For Fig. 7(a) and (b), the x-axis is the number of A-line, and the y-axis is the number of phase difference curves, that is time. The color from yellow to blue indicates the positive or negative change of the phase difference amplitude, and the lightness indicates the value of the phase difference amplitude. The position from the 1^st to 250^th line in Fig. 7(a) is the stationary part of the RNFL, the motion interference of which is eliminated by the mentioned phase extraction method. It can be seen from the figure that there is almost no significant phase difference change in the RNFL from the 1^st to 250^th line. The position marked by the red line is the 250^th to the 300^th line, which is the moving end of the RNFL. Obvious alternating light and dark changes can be seen in the figure, which means that the amplitude of the phase difference of the moving end of the RNFL has obvious changes over time. The change in the magnitude of the phase difference reflects the tissue movement of this part within 5s. In order to further quantify the movement of the RNFL, the phase difference data between the 250^th and 300^th line is extracted and shown in Fig. 7(b), then the data is averaged along the x-axis direction to get the curve of average phase difference versus time shown in Fig. 7(c). In order to compare the phase difference amplitude, change and the position of the peak and valley of the average curve intuitively, the x-axis of the graph is set to amplitude, and the y-axis is set to the number of curves. The averaging process is to reduce the jump error caused by the histogram extraction or phase compensation step in a single data. If only one line of the data is selected for analysis, one curve may contain the error caused by the histogram extraction algorithm, which will affect the overall pulsation curve extraction.

 

Fig. 7. The phase difference curves. (a) Phase difference change between 1st-300th A-lines; (b) Phase difference change between 205^th-300^th A-lines; (c) Phase difference curve after averaging.

Download Full Size | PDF

It can be seen from Fig. 7(c) that the changes of the phase difference amplitude correspond to the peak and valley changes of the averaged phase difference curve, indicating that the averaged phase difference curve can represent the displacement of the RNFL moving part. In order to further analyze the movement of RNFL, the x-axis and y-axis of the phase difference curve in Fig. 7(c) are exchanged and re-drawn as shown in Fig. 8(a), which can help us to analyses the RNFL movement over time more intuitively.

 

Fig. 8. (a) RNFL pulse waveform within 2s; (b) Cardiac pulse waveform within 5s.

Download Full Size | PDF

It can be seen from Fig. 8(a) that when the volunteers’ state is stable, the regular periodic waveform can be extracted. The RNFL motion curve shows a regular pulsating waveform within the 2s, and there are three complete pulsating waveforms, each cycle shows an obvious main wave. Compared with the cardiac pulse waveform within 5 seconds synchronously measured shown in Fig. 8(b), which also contains three complete similar cycles within 2 seconds, and have the same frequency and shape with RNFL pulse waveform. The measured heart rate is 91 pulses/min.

In order to further verify the nature of the extracted RNFL pulse waveform, analyze the relationship between the RNFL pulse waveform and the cardiac pulse waveform, Fourier transform was performed on the RNFL pulse waveform in Fig. 8(a) and (b) to analyze the two waveforms in the frequency domain, the result is shown in Fig. 9.

 

Fig. 9. RNFL pulse waveform and cardiac pulse waveform frequency domain diagram. (a) Spectrogram of RNFL pulse waveform; (b) Spectrogram of cardiac pulse waveform.

Download Full Size | PDF

In Fig. 9(a), the frequency spectrogram of the measured RNFL pulse waveform of motion end through Fourier transform contains a frequency of 1.27 Hz. In Fig. 9(b), the main frequency of the cardiac pulse wave after Fourier transform is 1.3 Hz. It is preliminarily proved that the RNFL pulse wave has the same frequency with the cardiac pulse wave measured at the same time.

4. Conclusion

To demonstrate the OCT could be used to extract the subtle RNFL motion using pulsation extraction algorithm in normal eyes, subjects’ retinal were imaged using SD-OCT with histogram-based RNFL pulsation extraction algorithm, which can be used for evaluating dynamic pulsatile characteristics of the nerve fiber layer in the human eye in vivo. Firstly, a series of OCT B-scan cross sectional images, were acquired with multiple retinal layers in normal subjects. The RNFL was identified from those B-scans. Secondly, based on the acquired cross-sectional OCT images, segmentation algorithm based on shortest path and convolutional neural network was applied to realize the segmentation between the outer membrane and the inner bound membrane, then a histogram-based RNFL pulsation extraction algorithm is proposed to track the displacement of the RNFL. Finally, evaluating the dynamic pulse signal extracted from the pulsating motion of RNFL. An infrared pulse sensor device is designed to collect cardiac pulse. The cardiac pulse waveform and the RNFL pulse waveform are compared and analyzed in time and frequency domain. The results validate that the RNFL pulse has the same frequency with the cardiac pulse.

5. Discussion

Here, a lot of research work has been carried out to study the pulsation model of retinal nerve fiber layer. Although some progress results have been achieved in the pulsation extraction of RNFL, there are still many problems that need to be improved. The extraction of the pulsation model in this thesis was performed based on healthy human eyes, which are characterized as motion with the heartbeat cycle and may have some delay. To further investigate the relationship between RNFL pulsation and heartbeat in time and amplitude, the next work can be performed by simultaneous acquisition of RNFL in healthy human eyes under different motion states compared with heartbeat. To further illustrate the correlation between glaucoma disease and RNFL pulsation, subsequent work could attempt to collect data from glaucoma patients for comparative analysis.

Driven by retinal micro vessels or choriocapillaris, impulses may not be important. Our research group has initially formed and we need more time to realize the significance of our findings in a comprehensive way. Our objective was primarily to report some novel evidence indicating that deformations occurred at the central and peripheral area of the ONH in glaucoma. The mechanism is of secondary importance and needs more focused efforts. In summary, the evidence we described above suggests ocular pulses have a plausible association with the alterations of pulsatile RNFL, which may make the amplitude more visible in glaucoma.

To determine the underlying mechanisms, the aim of our future research is to establish the pathophysiology models of such movements. We have developed a pulsatile RNFL extraction method based on OCT imaging system. To our knowledge, this is the first time that the pulsatile RNFL has been imaged and quantified. Our model especially excels at predicting the aberrations of the retinal periphery. In conclusion, we demonstrate that our high-speed SD-OCT system quantitatively measures pulsatile axial movements of the RNFL in human subjects. We have shown that the pulsatile RNFL movement is correlated with the pulsatile blood flow in the nearby large vessel. Pulse-induced RNFL movement is determined by biomechanical properties subject to characterization by this SD-OCT approach. Measurement of these properties may be clinically valuable in monitoring changes in such properties that may occur as RNFL damage progresses.