Optical coherence tomography (OCT) has become a standard diagnostic imaging tool in ophthalmology and is used in many other fields of medicine and biology. Whereas most OCT systems used in eye clinics are based on spectral domain OCT, swept-source OCT (SS-OCT) has particular advantages with respect to imaging speed and depth ranging [1–3]. Such high-end swept-source OCT devices provide exquisite resolution and their speed supports imaging with a large field of view. Furthermore, modern swept sources allow for full coverage of the human eye, from cornea to retina, without a critical loss of sensitivity. Moreover, since no bulky high-resolution spectrometer is needed, SS-OCT technology allows for a potentially smaller system footprint. However, these systems are still expensive, mainly due to the price of the swept source itself, since the complexity of achieving stable tuning over a broad spectral bandwidth whilst maintaining good coherence remains high and costly. Owing to the need for large-scale accessibility of ophthalmic OCT systems, we are currently witnessing a trend for developing miniature, low-cost, or homecare/point-of-care OCT devices which would enable regular screenings [4,5], even in low-resource settings.

Distributed feedback (DFB) and distributed Bragg reflector (DBR) lasers might be candidate sources; however, they achieve small thermal tuning ranges and exhibit critical mode behavior during tuning. Recently, thermally tuned vertical-cavity surface-emitting lasers (VCSELs) have been proposed as a low-cost alternative for SS-OCT. VCSELs with mechanical cavity length tuning are already commonly used in OCT [3], but they are still expensive due to their complexity. First approaches using thermally tuned VCSELs that were centered on telecom wavelengths, such as 1300 nm [6], achieved a tuning speed of 100 kHz. These types of lasers excel at instantaneous linewidth, resulting in coherence lengths of several centimeters, which are mainly limited by the detection bandwidth. Nowadays, VCSELs are most commonly designed to have a center wavelength of 850 nm. The lower wavelength is better suited for retinal imaging, as well as biometry, due to the lower water absorption compared to 1300 nm. As a consequence of their application in data centers, smart phones, as well as lidars, they are produced in very large volumes and are hence several magnitudes cheaper than high-performance swept sources specifically designed for OCT.

In the following, we demonstrate first in vivo results for full-eye biometry and anterior segment SS-OCT with a thermally tuned VCSEL at 850 nm (TRUMPF ULM-850-B2) used as a low-cost swept source. Sweeping was achieved through current modulation, resulting in thermal tuning of the output wavelength. The source was fiber pigtailed and delivered an output power ex-fiber of 1 mW at a wavelength of 850 nm. The tuning bandwidth depends on the operating temperature as well as on the tuning speed. In our case, the VCSEL was operated at room temperature without temperature control. The current driving signals are shown in Fig. 1(a). The modulation in use was a 2 kHz sawtooth of 0–16 mA with a duty cycle of 50%, allowing sufficient time for the source to cool down in between the sweeps. For all tested tuning-speed settings, the wavelength of the instantaneous output of the laser always returned to baseline. Hence, there was no visible drift of the wavelength during operation that could be caused by insufficient thermal relaxation and the accumulation of heat over time.

 

Fig. 1. Source characteristics: (a) VCSEL driving settings with resulting spectra; (b) collimated beam profile ex-window; (c) VCSEL spectral envelopes at 2, 5, 10, 20, 50, and 100 kHz.

Download Full Size | PDF

The dependence of the tuning bandwidth on the speed was assessed with a high-speed spectrum analyzer (Ando AQ-6315E). The resulting spectral sweep performance of the VCSEL is plotted in Fig. 1(c), which shows the integrated spectrum over repeated spectral sweeps at 2 kHz, 5 kHz, 10 kHz, 20 kHz, 50 kHz, and 100 kHz, respectively. The peaks in the spectra at the edges result from a slower tuning speed at those wavelengths. The spectral envelopes allow us to observe the bandwidth as a function of sweep rate. The lower bandwidth in spectra with larger sweep rates is assumed to be due to insufficient heating and cooling of the VCSEL’s internal temperature during the sweep. The source output profile ex-window is shown in Fig. 1(b). The strong side lobes resulted in a power loss of about 40%, when coupling the light into a single-mode fiber.

In order to assess the OCT performance, we set up the system shown in detail in Fig. 2. Due to the non-linear behavior of the VCSEL’s spectral tuning phase, part of the light from the source was sent to a reference Mach–Zehnder interferometer (MZI) using a 95/5 fiber splitter. The phase of the reference interferogram at a fixed delay Δz_MZI was then used to linearize the recorded spectra with respect to the wavenumber k. The Mach–Zehnder-type OCT interferometer consisted of a 50/50 fiber coupler to split the light into reference and sample beams. Polarization control paddles in the reference arm allowed the polarization of the returning sample arm light at the detector to be matched. In the sample arm, the light was coupled out via a fiber collimator (beam diameter: 2.4 mm) and focused by an achromat (100 mm) onto the sample. For ocular imaging, a fixation light was coupled into the beam path by means of a cold dichroic mirror (740 nm cutoff wavelength). Both the reference and the OCT interferometric information was acquired using dual-balanced detectors (OCT: Thorlabs PDB450A-AC, reference: Thorlabs PDB460C-AC). The OCT detector amplification was set to 10⁵, with a detection bandwidth of 4 MHz. Synchronization between the VCSEL driver sweep and data acquisition was achieved using an electronic A-scan trigger signal generated by a field-programmable gate array (FPGA). Data from both detectors were acquired in parallel using a waveform digitizer (Alazartech ATS9360) with a sampling rate of 10 MSps. When 25 kHz and 50 kHz sweep rates were used, the sampling rate was increased to 250 MSps.

 

Fig. 2. System setup schematics consisting of two parts: the reference interferometer and the Mach–Zehnder-based OCT setup. FC: fiber coupler, FCl: fiber collimator; L: lens, DM: dichroic mirror, FS: fixation screen; MS: MEMS scanner (optional); DBD: dual-balanced photodetector.

Download Full Size | PDF

The experimental performance parameters are given in Table 1. At 2 kHz tuning rate, the system achieved a sensitivity of 97 dB with a sample beam power of 0.3 mW and an axial resolution of 50 µm in air. The lower sample arm power for 25 and 50 kHz was a result of adding the scanners and relay optics.

Table 1. Settings and Results Obtained at Different Sweeping Rates

View Table

Another important strength of VCSELs is their long coherence lengths. According to the manufacturer, the nominal instantaneous linewidth is 100 Hz, corresponding to a coherence length of about 3 m in air for continuous wave operation. Figure 3 plots the sensitivity roll-off in depth in pulsed operation mode for tuning rates of 2 kHz, 25 kHz, and 50 kHz, respectively. At 2 kHz, a low roll-off of −3 dB at most was measured with a 5 cm round trip delay. The VCSEL is therefore well suited for long-range interferometric distance sensing, including ocular biometry. The stronger roll-off for 25 kHz and 50 kHz was due to the limited electronic detection bandwidth of 45 MHz at the detector amplification setting used (10⁴).

 

Fig. 3. Signal roll-off at A-scan rates of (a) 2 kHz, (b) 25 kHz, and (c) 50 kHz. (d) Comparison of signal roll-off for A-scan rates of 2 kHz, 25 kHz, and 50 kHz.

Download Full Size | PDF

Thermal sweeping of the VCSEL exhibited slight temporal fluctuations of the spectral range. Even after correcting for the non-linear sweep, i.e., k-mapping, the spectral shifts caused fluctuations of the imaging range after the Fourier transform.

This was due to the intrinsic relation between the spectral sweep parameters and the depth range z_max given as ${z_{max}} = \lambda _0^2/({4\delta \lambda } )$, where $\delta \lambda $ is the spectral sampling interval and ${\lambda _0}$ is the center wavelength. In order to demonstrate this effect, a calibration test target with two distinct interface signals with a distance of 3.22 mm was used as a sample for recording 1000 spectra at 2 kHz sweep rate. Figure 4(a) shows the reconstructed A-scans after k-mapping and 5-fold zero padding of the spectra.

 

Fig. 4. Thermal tuning fluctuation correction: upper row shows plots of a series of 1000 A-scans with corresponding M-scans displayed in the lower row (a) after k-mapping, (b) after axial registration to the first peak, and (c) after tuning fluctuation correction.

Download Full Size | PDF

The spectral fluctuations resulted in differently axially scaled A-scans in time. The indices of the first and second peak maxima had standard deviations of 19 and 23 pixels, respectively. The distance in between varied with a standard deviation of 4.2 pixels. Even after the registration of the A-scans with respect to the first interface using cross correlation, the position of the second interface varied visibly [Figure 4(b)]. In order to correct for the A-scan scaling, we determined the actual depth range for each sweep by using the known optical path length of the reference interferometer $\Delta {z_{MZI}}$. Subsequently, we determined the shortest depth range of the series of recorded A-scans and scaled all other scans using linear interpolation. Formally, if $z_{MZI,p}^{(i )}$ is the pixel count of the peak position after applying the Fourier transform to the ith k-linearized MZI reference signal, the corresponding full depth range for both the reference and the OCT A-scan signals is

To demonstrate the system performance in vivo, an eye of a healthy volunteer was imaged. With the focus set to the anterior chamber, we obtained signal peaks for all eye structures relevant for determining intraocular distances.

Figure 5(b) shows the average of 1000 A-scans recorded in 500 ms, where prominent signal peaks for the cornea, lens, and retina can be observed. The corresponding M-scan consisting of 1000 A-scans at the same location is presented in Fig. 5(c). Correction of the spectral tuning fluctuation and axial registration with respect to the corneal signal have been applied. Measurement of the axial length (AL) requires the determination of the exact positions of the corneal front surface and the retinal interface. As a proof of concept, the corresponding peaks were extracted using the peak detection algorithm findpeaks() in MATLAB. The AL extraction precision was determined by first extracting the ALs from averages of 100 A-scans and then calculating the standard deviation over 10 such ALs. The average optical AL was 32.4 mm with a standard deviation of 24 µm, which was better than the axial resolution of the system. Following Haigis [7], the geometrical AL was determined as 23.62 mm, which is close to the ground truth of 23.66 mm obtained by a commercial biometer (IOL Master 500, ZEISS, Jena, Germany).

 

Fig. 5. Ocular interfaces: (a) schematic; (b) average A-scan; (c) M-scan (consisting of 1000 consecutive A-scans).

Download Full Size | PDF

The integration of MEMS scanners into the OCT optical system enabled the acquisition of full-eye scans across a depth range of 5 cm and a lateral range at the cornea of 11 mm. The A-scan rate for these images was 25 kHz and 50 kHz, respectively, with a power on the sample of 0.21 mW. The sampling rate was increased to 250 Msps. The use of depth enhancement by flattening the noise floor through the application of a linearly scaled multiplicative factor with depth resulted in a more homogeneous contrast across the whole depth range and facilitated the recovery of retinal peaks [Figs. 6(a)–6(c)]. The decrease in performance at the 50 kHz compared to the 25 kHz A-scan rate is due to the decline in tuning range and related axial resolution as well as the stronger roll-off (cf. Fig. 3). Note that the detection bandwidth for OCT recording was narrower than that for recording the roll-off curves to allow a higher sensitivity closer to the zero delay.

 

Fig. 6. (a) Impact of depth enhancement on A-scans. (b) Original full-eye scan at 25 kHz A-scan rate. (c) Depth-enhanced full-eye scan at 25 kHz A-scan rate. (d) Depth-enhanced full-eye scan at 50 kHz A-scan rate. 40× B-scan averaging was applied for all scans. White arrows point at the retina.

Download Full Size | PDF

In summary, we have developed a low-cost swept-source optical coherence tomograph (SS-OCT) using a thermally tuned vertical-cavity surface-emitting laser (VCSEL). Sweeping at an A-scan rate of 2 kHz resulted in a system sensitivity of 97 dB and an axial resolution of about 50 µm in air. Intrinsic fluctuations of the tuning wavelength range were corrected for in post-processing using a reference interferometer signal. The system characteristics proved adequate for performing optical biometry on healthy volunteers. Driving the source at 25 kHz allowed for in vivo full-eye OCT scans. Such scans allow the extraction of important biometry parameters in addition to the AL, such as the corneal curvature or the lens position and orientation. Those are of interest, for example, for estimating the post-refractive-surgery outcome. Furthermore, with the retinal structure in a tomogram at hand, it is easier to determine the AL with higher precision due to the visible retinal reference interface which distinguishes the inner limiting membrane from the retinal pigment epithelium. Pushing the A-scan rate to 50 kHz helps to reduce the measurement times for OCT, but comes already at further reduced axial resolution.

The current drawbacks of the low-cost system are its low optical output power and limited tuning bandwidth. The low optical power of the VCSEL could be mitigated by using semiconductor optical amplifiers; however, this might critically increase the price of the system. Another possible solution might include a combination of several VCSELs, as is commonly the case with LEDs, in order to enhance the luminance. Bandwidth could be increased by combining the VCSELs, as is done for superluminescent diodes. Nonetheless, as stated in the work of Harper and Vakoc [8], a lower axial resolution results in a higher signal-to-noise ratio (SNR) for diffusely scattering samples. Furthermore, there is a current trend for developing homecare and point-of-care OCT devices. These might trade-off image quality for system costs, provided that the image still allows the recognition of signs of pathologies. e.g., open-angle glaucoma or retinal pathologies [9].