Development of next generation low-cost OCT towards improved point-of-care retinal imaging

Abstract

Low-cost optical coherence tomography (OCT) has shown promise in increasing access to noninvasive retinal imaging at the point of care, especially in low-resource environments. A next-generation low-cost OCT system is presented which improves performance over previous versions by employing balanced detection, improved spectrometer falloff, and an increased A-line rate of 40 kHz. An algorithm is presented for image display that uses a histogram matching procedure to improve contrast-to-noise ratio (CNR). Imaging performance is benchmarked with CNR analysis of retinal OCT images, demonstrating a CNR of 2.01 ± 0.39 (p < 0.0001) for macula images collected during a clinical trial, a significant improvement over previous low-cost OCT systems.

1. Introduction

Optical coherence tomography (OCT) is a non-invasive, high-resolution optical imaging technique that enables cross-sectional imaging of biological tissues. Given the optical transparency of ocular tissues, OCT has been widely adopted by ophthalmologists as the gold standard for non-invasive examination of retinal pathology [1–3]. Modern full-featured commercial OCT systems include eye tracking, automatic pupil alignment, fundus imaging, and automated focus and reference arm adjustment to improve ease of use. However, these full-featured clinical systems can cost anywhere from $\mathrm{\$}40\text{k}$ to $\mathrm{\$}150\text{k}$ [4], limiting access to large ophthalmology clinics and hospital networks.

To enable wider access to OCT imaging, several approaches have been advanced. The first clinical implementation of low-cost OCT [5] was realized by leveraging advances in 3D printing, combined with custom electronics and software. These systems are shoebox-sized [4,6] and can be easily carried between patients at the point of care. Low-cost OCT systems for research purposes have been commercialized and can be purchased for $\mathrm{\$}12\text{k}$ to $\mathrm{\$}20\text{k}$ . Other approaches for increasing access include development of briefcase-based OCT systems [7] and integration of handheld probes [8,9]. Chip-based OCT or photonic integrated circuits [10,11] promise to reduce the physical footprint of OCT systems even further. However, these efforts are still in development and have not yet realized an OCT system suitable for use as a diagnostic tool in the clinic. Home-based, self-guided OCT [12] has been developed with the goal of helping physicians track age-related macular degeneration (AMD) progression between clinic visits without the need for a skilled imaging technician. Notal Vision Home OCT is in ongoing clinical trials and has recently received FDA De Novo authorization [13]. This is a promising example of the momentum behind broader access to retinal diagnostics. However, this system is highly specialized for monitoring fluid in AMD and is not intended for use in general retinal diagnostics in low-resource environments.

Previous low-cost OCT systems sacrificed image quality or features to reduce cost compared to full-featured commercial devices. Our first-generation low-cost OCT system (Gen 1) achieved good performance but was still significantly less than that of the Heidelberg Spectralis (Heidelberg Engineering, Heidelberg, Germany), with the difference attributed to slower scan speeds and low scanner output power [5]. Costs were reduced by using superluminescent diodes (SLD) without a controlled thermoelectric cooler (TEC) and a custom spectrometer made from a 3D printed housing, but both of these components impacted system performance.

In this work, we present two next generation low-cost OCT systems that advance the performance, features, and image quality over the Gen 1 system. The Gen 2 system [14] introduces hardware modifications to stabilize SLD output power and reference arm power, increase A-line rate, and improve operability. In addition, a self-referencing histogram matching procedure has been developed to enhance contrast through post-processing. Further improvements were made with Gen 3, including balanced detection to further improve image quality. While balanced detection in spectral domain OCT reduces common mode noise, it requires the use of a second spectrometer [15–19]. However, in the low-cost implementation, this is a relatively small cost increase. This work compares the design and operating characteristics of both Gen 2 and Gen 3 at the system level and benchmarks their imaging performance against the Gen 1 system and a commercial OCT system via clinical imaging of patient retinas.

2. Methods

The progression of three generations of low-cost OCT systems is presented in Fig. 1. The topology of these systems is unique in that the scanner contains the SLD light source, fiber-based interferometer, and control electronics. Thus, only a single fiber optic cable is returned to the engine for detection by the spectrometer. The engine also contains a computer for instrument control, image processing and display, as well as electrical power distribution circuits. The major changes across the generations include scanner and interferometer design, spectrometer design and operability features. These changes are summarized in Table 1. All three generations used Exalos SLDs (Exalos AG, Schlieren, Switzerland) as light sources. The operating computer was a small form factor PC for Gen 1 and Gen 2 (Intel NUC), while a separate laptop (Lenovo ThinkPad) was chosen for Gen 3. The spectrometer sensor was upgraded from Gen 1, which used a Hamamatsu detector (Hamamatsu, Hamamatsu City, Japan), to Teledyne e2 v line scan cameras (Chelmsford, UK) in Gen 2 and Gen 3. Changes in system components and features also resulted in different costs for each system; The Gen 1 system cost $\mathrm{\$}5037$ [5], while the Gen 2 system cost $\mathrm{\$}6097$ , primarily due to adoption of the faster camera. The Gen 3 system also increases cost to improve performance with the additional spectrometer adding approximately $\mathrm{\$}2500$ (includes a second, upgraded 40 kHz sensor) with another $\mathrm{\$}1500$ increase due to the use of the higher performance laptop PC.

Fig. 1.

Low-cost OCT design evolution. Example photos show the scanner and engine of each generation.

Table 1. Summary of changes between low-cost OCT Generations.

	Gen 1	Gen 2	Gen 3
Interferometer	50:50 Michelson	50:50 Michelson	Mach-Zehnder
SLD Cooling	Not Controlled	TEC-Cooled SLD	TEC-Cooled SLD
Pupil Camera	No Pupil Camera	Pupil Camera	Pupil Camera
A-line Rate	12.5 kHz	40 kHz	40 kHz
Display	7” Touchscreen	10” Touchscreen	Laptop Display
Detection Format	Single	Single	Balanced
Processor	Intel i5 (2 Core, 1.6 GHz)	Intel i7 (4 Core, 2.7 GHz)	Intel i7 (14 Core, 3.5 GHz)
Scanner Body	Handheld	Fixed	Fixed
System Cost	$5037	$6097	$9522

2.1. Scanner optical power output

To reduce cost, the Gen 1 low-cost OCT system used an SLD without a TEC. However, this open-loop cooling approach left the Gen 1 system susceptible to optical power fluctuations, which is not ideal for clinical imaging. Specifically, scanner output power varied between 400 µW and 680 µW over 20 minutes of imaging [5,14], which significantly limited image quality. For Gen 2, the current controller for the SLD was upgraded to include a TEC driver to stabilize the output power between 660 µW and 680 µW [14]. This custom controller was also used in the Gen 3 system, which produced stable power levels of 720 - 730 µW. All three systems used SLD’s with the same center wavelength of 840 nm and with 42-47 nm FWHM bandwidth.

2.2. Scanner design

All three systems used off the shelf optics and 3D printed (Form 3, Formlabs, Boston, MA) optical mounts to relay light to and from the patient’s retina. Micro-electro-mechanical system (MEMS) mirrors are used in all three systems for optical scanning and liquid lenses are used for dynamic focus control. Scanner optics for all three generations were selected using OpticStudio (Zemax, LLC. Kirkland, WA). The Gen 1 and Gen 2 systems used Michelson interferometers consisting of 2 × 2 50:50 fiber-optic couplers (FC). The first output was directed towards the sample arm and the other was directed towards a reference arm consisting of lens tube and a mirror, as depicted for Gen 2 in Fig. 2(a). Back reflected light from the sample and reference arms interferes within the fiber coupler and is directed to the engine where the interference signal is recorded by a spectrometer. A pupil camera was added to the Gen 2 system to assist the operator in scanner placement relative to the eye. The USB controlled pupil camera (ELP, Shenzen, China) was integrated into the optical train of the OCT scanner optics using a dichroic cold mirror. A tunable liquid lens (TL1, Optotune, Dietikon, Switzerland) was added to the fixed reference arm of the Gen 2 system which introduced a slight defocus to allow the returning power to be adjusted by the operator. This feature was incorporated to avoid acquiring saturated interference signals. The sample arm liquid lens (TL2) has been used since Gen 1 for manual focus control [5].

Fig. 2.

Diagram of scanner (left column) and engine (right column) components for Gen 2 (a) and Gen 3 (b) low-cost OCT systems. CM- cold mirror, CR- circulator, FC- fiber coupler, L- lens, M- mirror, PC- polarization controller, SLD- superluminescent diode, SR- spectrometer, TL- tunable lens.

The Gen 3 system was further upgraded to advance imaging performance and to improve operator control via feedback from a pupil camera and an adjustable eye fixation target (Fig. 2(b)). To improve signal quality, the Michelson interferometer used in Gen 1 and Gen 2 was replaced by a modified Mach-Zehnder interferometer including an optical circulator [20] to enable balanced detection. The pupil camera was upgraded to provide greater depth of focus and to improve utility. Finally, the Gen 3 system added an external fixation target that can be electronically controlled by the operator to allow imaging of various retinal features.

The modified Mach-Zehnder interferometer used in Gen 3 was based on a design suggested by Rollins and Izatt [20]. A 2 × 2 90:10 FC was used to split the light from the source with 90% going to an optical circulator in the sample arm and 10% going to the reference arm, as shown in Fig. 2(b). The output of the second port of the optical circulator was directed to the scanner, with the returned signal directed by the circulator to the third port. This signal and the light returning from the reference arm are combined at a 2 × 2 50:50 FC housed in the scanner enclosure. There is a π phase shift between each interference signal allowing for a differential measurement that increases signal recovery and reduces noise when compared to single channel detection [15–20]. In Gen 1 and Gen 2, the Michelson interferometer could only detect 50% of the interfered light. The new Gen 3 design now recovers nearly 100% of the signal.

The Gen 3 system also includes features to better aid the operator in aligning the scanner to the patient’s eye. The Gen 2 system added a pupil camera for a similar purpose but provided limited depth of focus, restricting utility. This was corrected in Gen 3, by providing a greater depth of focus. The Gen 3 system also implemented a low-cost external fixation target based on an LED matrix, driven by an Arduino microcontroller and a joystick for control. Using the joystick the operator can guide the patient’s off eye or fellow eye, i.e., the eye not being imaged, to different locations on the LED matrix. For most patients, the eye being imaged naturally tracks with the fellow eye, allowing the operator to have more control of the location on the retina images were acquired.

2.3. Spectrometers

Low-cost spectrometers, like the one used in the Gen 1 system, have been described previously [5,6]. Similar to Gen 1, the Gen 2 and 3 spectrometers were designed in OpticStudio with off-the-shelf optics; their housings were designed in SolidWorks and 3D printed (MakerBot MethodX, New York City, NY) using an ABS material that is robust to changes in temperature. The spectrometer design used in Gen 2 (Fig. 3(a)) is very similar to that in Gen 1, both used and off-axis parabolic (OAP) mirror to collimate and steer the input signal. To improve the optical throughput and simplify alignment, the Gen 3 spectrometer design was revised to use a collimating lens instead of the OAP. Spectrometer performance was evaluated in terms of sensitivity falloff [6,15].

Fig. 3.

Procedure for self-referenced histogram matching (HM) to enhance retinal signal. In each step the whole image (top) and retinal signal histogram (bottom) are displayed to show the incremental changes. First, a 3-pixel lateral median filter is applied to the original image. Next, the original image is histogram matched to the filtered image using a polynomial fit. Lastly, the black and white points of the histogram matched image are adjusted to achieve desirable contrast. Magnified views of each step extracted from the region highlighted by the dashed red box are shown on the bottom right.

2.4. Engine

The low-cost OCT system engine consists of the spectrometer, computer, and power control circuitry. The engine connects to the scanner via fiber optic and electronic connection. In addition to the spectrometers, many parts of the engine have changed over the three generations including the housing, interface, and computer as seen in Fig. 2. The Gen 1 engine housing was 3D printed [5,6] while the Gen 2 and 3 housings were made of sheet metal and 3D printed faces. For Gen 2, the touch screen interface was increased from 7” to 10” for easier operation of the device and a larger viewing area. In Gen 1 and 2, a small form factor computer (Intel NUC) was built into the engine and connected to the touch screen for display and control. In Gen 3 these were replaced with an external Windows computer connected by a USB port. This offered a more powerful GPU for increased processing load to better facilitate balanced detection. Software was upgraded to handle the dual data stream acquisition and subtraction for balanced detection.

2.5. Clinical study

The performance of the Gen 2 and Gen 3 system was evaluated with retinal images acquired in a clinical setting with IRB approval from Duke. The Gen 2 characterization study was described previously [14]. Briefly, the Gen 2 scans were acquired from 12 patients with no history of retinal disease. Horizontal and vertical scans were acquired from both eyes of each patient. The scans consisted of 25 repeated B-scans which were used to perform a range of 2-25 averages. The minimum of 2 was used for direct comparison with the Gen 1 study. The range of 25 was to establish the trend in CNR with frame average. For Gen 3 scans were acquired from 14 patients with no history of retinal disease (8 males and 6 females, an average age of 27.6 years old and a range of 23–32 years old). The acquired scans were centered at the macula, optic nerve head (ONH), and blood vessels surrounding the ONH for both eyes. A radial scan configuration was used, featuring 17 B-scans rotated about the same central point in an asterisk pattern. To eliminate operator bias in selecting scans for analysis, a modified quality index (QI) [21] was used to quantify scan quality. Acceptable images were defined to have a QI threshold of 35 for at least 8 of the 9 repeated images used for averaging.

2.6. OCT processing and Histogram matching

Standard SD-OCT processing steps were used to produce images from raw spectrometer data, including background subtraction, wavenumber interpolation, and dispersion compensation. Image registration was applied to compensate for subject movement between frames to allow alignment for frame averaging to reduce noise. Image contrast was adjusted for each image by examining the histogram of pixel intensities and introducing adjustments. In the Gen 1 study, histograms of the images from the low-cost OCT images were adjusted to match those of the same patient from the HE system [5]. This step was performed to allow for a fair comparison between the low-cost and commercial systems; histogram analysis showed that HE images had near-gaussian distributions while low-cost images resembled Rayleigh distributions. This approach, however, limits the usefulness of the method to cases where a reference image from the same patient is available.

An innovation of the Gen 2 system was the introduction of a self-referenced histogram matching (HM) procedure which helps to standardize the layer contrast across different scans while also improving contrast-to-noise (CNR) [14]. Figure 3 shows the flow chart for our HM procedure. Images from the HE system had an average black level set to about 30% of the standard image which was used as a target for our HM procedure. To provide a Gaussian distribution model for the HM procedure without a reference image, a 3-pixel lateral blurring was used. Although the blurred image has reduced image resolution, it provides a suitable model histogram for adjusting the original, unblurred image. Thus, better CNR is obtained without sacrificing image resolution.

The analysis of Gen 2 images also showed a high concentration of noise in the vitreous region with the degree of noise varying with image depth. To compensate, a 2^nd depth-dependent background subtraction was implemented. A-lines from background scans were used to fit a 3^rd degree polynomial of intensity as a function of depth and subtracted from the unblurred image. The background subtracted image was histogram matched to the blurred image for the final image output. To further enhance CNR while minimizing loss of retinal signal, the histogram matching procedure was altered. The background subtracted image, instead of the original image, was matched to the blurred image.

To accommodate the new balanced detection scheme, the OCT processing for Gen 3 was modified. A pixel map from SR1 to SR2 was computed from a background scan. The SR1 signal was then resampled and rescaled in the coordinates of SR2 before performing balanced subtraction. Next standard SD-OCT processing steps of wavenumber interpolation, dispersion compensation, registration, and averaging were performed. The self-referencing histogram matching procedure was adapted to the balanced images and applied before measuring CNR. The segmented retinal signal from HM images was used as the signal portion while the noise region was defined as a 10 × 10 pixel region above the RNFL. CNR was calculated via Eq. (1) where s is signal, and n is noise.

$$CNR=\frac{\overline{s}-\overline{n}}{\sqrt{{\sigma }_s+{\sigma }_n}}$$ (1)

3. Results

3.1. Spectrometer characterization

To determine the OCT sensitivity fall-off with depth, an alignment interferometer was used. The optical path length difference was adjusted in increments of 0.15 mm from 0 mm to 3 mm using micrometer translation stages. The locations of the peaks for each measured depth were fitted to a 2^nd order polynomial to find the z_6 dB depth. SR1 and SR2 were both measured to have z_6 dB of 2.1 mm (Fig. 4(a)). After remapping the SR2 to the coordinates of SR1 and computing the balanced signal, the z_6 dB was measured to be 1.9 mm. While not as deep as the single detection z_6 dB, the balanced detection measurements have an increased signal amplitude of ∼6 dB (Fig. 4(a)), resulting in a greater signal at depth. The similarity between the two SRs is quantified by the pixel offset and tilt (Fig. 4(b)). The measured slope was 1.02 pix_SR2/pix_SR1, offset equal to -8.62 pix_SR2, and R² of 1.00.

Fig. 4.

Spectrometer characterization for Gen 3 system. (a) The individual performance of SR1 (red), SR2 (blue), and balanced detection (black) respectively, over 3 mm of depth. The points represent individual measurements, and the dashed lines represent the fitted line for calculating z_6 db. (b) The alignment offset between SR1 and SR2. (c-e) Example macula image acquired with Gen 3 low-cost OCT showing the change in image quality when using (c) single detection, (d) balanced detection, and (e) balanced detection with self-referencing histogram matching. The measured CNR is shown for each image. 9 B-scans used for averaged images. Scale bars: 500 µm.

3.2. Image quality

Example retinal images are shown in Fig. 4(c-e) when applying single detection, balanced detection, and histogram matched balanced detection. Each example is a macula scan collected from a patient using the Gen 3 system. The single detection (Fig. 4(c)) image is dimmer than the balanced image (Fig. 4(d)) across all depicted retinal layers when displayed with same contrast values. This makes sense considering the 6 dB improvement in recovered signal with balanced detection. The measured CNR also increases from 1.06 to 1.26, a 19% improvement. After histogram matching (Fig. 4(e)), the CNR improves to 1.83, a 73% improvement from single detection and 45% improvement from balanced detection without histogram matching. The choroidal structure is more evident after HM (Fig. 4(e)) than with balanced detection alone (Fig. 4(d)). Being able to see deeper structure can aid physicians in assessment of retinal features.

Additional retinal images acquired with the Gen 3 system are shown in Fig. 5, which highlight the system’s improved image quality across the eye. The external fixation target enables visualization of other retina regions including macula (Fig. 5(a)), optic nerve (Fig. 5(b)) and the vasculature superior to the optic nerve (Fig. 5(c)).

Fig. 5.

Example Gen 3 retinal images from (a) macula, (b) optic nerve, and (c) vasculature superior to the optic nerve. 9 B-scans used for averaged images. Scale bars: horizontal = 500 µm, vertical = 250 µm.

3.3. CNR comparison

To evaluate the performance of the next generation low-cost OCT systems, we compared the CNR B-scan images from healthy subjects for the Gen 1, Gen 2, and Gen 3 systems against those from the HE. First, each system was individually compared against the HE. Then, a Welsch’s ANOVA with non-homogeneous variances followed by a Games-Howell post-hoc test (JMP Pro) was used to check for significance against the other systems.

The HE and Gen 1 analysis comprised 99 measurements from 25 patients [5] resulting in mean CNR of 1.80 ± 0.33 and 1.69 ± 0.27 (p = 0.06) respectively [14], for averages of 2 B-scans. A paired t-test showed that the 6% difference was statistically significant. Gen 2 was characterized using 40 measurements from 12 patients with a mean CNR of 1.74 ± 0.07 for 2 B-scan averages. A two-sample t-test with unequal variances showed no significant difference in CNR between Gen 2 and HE (3% difference, p = 0.32). The comparison between Gen 1 and Gen 2 also did not show a statistically significant difference. A one-way ANOVA across the CNR values for the HE, Gen 1, and Gen 2 systems agrees with the significance of the previous direct pairwise comparisons.

To analyze the change in CNR with the number averages, the mean CNR for Gen 2 was characterized for various numbers of scan averages (N), as shown in Fig. 6. For comparison the averages for N = 9 and 25 are seen to increase to 1.93 ± 0.08 and 2.03 ± 0.1, respectively. As shown above in Eq. (1), CNR can be treated as the square of the difference between the signal-to-noise ratios (SNR) of the signal and noise regions of the image. Upon averaging N B-scans with independent noise contributions, CNR is expected to increase by a factor of $\sqrt{N}$ . Upon analyzing our results (Fig. 6) we see that the relation between CNR and N is well modeled by Eq. (2), where a, b, and c are fitted coefficients.

$$CNR(N)=a{e}^{bN}\sqrt{N}+c$$ (2)

Fig. 6.

Measured CNR (points) and fitted CNR (black line) as a function of number of B-scan averages for Gen 2 system. Measured values lie within the 95% confidence interval (shaded region) for fitted coefficients in Eq. (2): a = 0.147 (0.138, 0.157), b = -0.0177 (-0.0191, -0.0163), c = 1.55 (1.536, 1.573), R² = 0.997. Error bars give the standard deviation in measured CNR.

The Gen 3 study was characterized using 310 macula measurements from 13 patients, after filtering for QI above 35. The mean CNR for Gen 3 was found to be 2.01 ± 0.39, a statistically significant (p < 0.0001) increase of 12% over the CNR for the HE result. Similarly, the CNR for the Gen 3 images is significantly greater (p < 0.0001) than for the other low-cost systems. A one-way ANOVA shows that the differences between the other systems are no longer significant (p > 0. 05) with Gen 3 offering superior performance across the group. These results are summarized in Table 2. Figure 7 illustrates the distribution of healthy macula CNR measurements for all systems.

Table 2. Results of CNR across all Systems.

System	Against	Diff CNR	Std Err Diff	p-Value
Gen 3	Gen 1	0.319	0.025	<.0001
Gen 3	Gen 2	0.270	0.017	<.0001
Gen 3	HE	0.210	0.028	<.0001
HE	Gen 1	0.109	0.031	0.0603
HE	Gen 2	0.060	0.025	0.3276
Gen 2	Gen 1	0.049	0.021	0.3412

Fig. 7.

Distribution of CNR values for macular images across 3 Generations of low-cost OCT systems. *** - p < 0.0001

4. Discussion

The Gen 3 low-cost OCT system significantly outperforms earlier generation systems in terms of CNR while also offering design and software improvements. The better performance is enabled by increased spectrometer falloff, more efficient interferometer design, balanced detection, and faster data acquisition rate. Gen 3 also enables control of subject eye fixation via an adjustable external target. This expands the scannable region beyond just the macula and enables access to a greater range of retinal diagnostics. Even though the Gen 2 system included some of these upgrades, such as the 40 kHz line rate, the integrated TEC driver, pupil camera, and scanner optics configuration, the additional improvements of the Gen 3 system were needed to realize a superior level of performance.

A key reason for the better performance of the Gen 3 system is the change in interferometer topology. As mentioned in section 2.2, the modified Mach-Zehnder interferometer with optical circulator permits recovery of nearly 100% of the interference signal instead of the 50% recorded with the Michelson interferometers used in Gen 1 and Gen 2. Further, the Gen 3 spectrometer was re-designed to be more efficient and more robust than Gen 2. The OAP collimation used for Gen 2 offered some advantages in terms of cost, but it was found to be sensitive to changes in position and orientation. Tolerance analysis showed this could lead to a 70% power loss. The effects of this improved efficiency can be seen in Fig. 5(e), where both SR1 and SR2 start out with dynamic range of about 114 dB, an improvement over previous low-cost systems which provided approximately 100 dB, realizing a remarkable signal level for spectrometer-based OCT. However, using balanced detection provides even greater dynamic range of almost 120 dB, rivaling some swept-source OCT systems [22,23].

The Gen 3 system also includes usability features that allow the operator to more quickly and efficiently acquire OCT images. The upgraded pupil camera enables centration on the patient pupil, a feature that was lacking in earlier generations. The fixation target allows patient gaze to be directed to variable locations. This enables imaging of the ONH, for example, which is of particular interest for diagnosing various pathologies such as glaucoma. This feature offers potential benefit for neurology where OCT is being explored for detecting neurodegenerative diseases like Alzheimer’s disease [24–27]. Here the improved ease of alignment may offer benefit in examining older patients, a benefit we will seek to establish in future studies. Further to the point of ease of use in this patient population, we chose to implement Gen 3 as a fixed scanner. Although, the Gen 1 scanner body was designed to allow for handheld use, alignment with the patient proved to be more challenging in handheld mode than fixed mode, even though no difference in CNR was seen between the two modes [5]. Thus, Gen 3 is designed solely to operate as a fixed scanner.

A significant tradeoff for the improved performance of the Gen 2 and Gen 3 systems compared to Gen 1 was cost. While these systems are still far less expensive than commercial systems, the Gen 2 system costs $1060 more than Gen 1. Most of this increase is due to upgrading the spectrometer sensor from 12.5 kHz to 40 kHz A-line rate. The Gen 3 system costs $1900 more than the Gen 2 system due to the addition of the second 40 kHz spectrometer, as well as the optical circulator and fixation target. This increased cost however does result in a significantly higher performing low-cost OCT system. The laptop was chosen instead of the internal small form factor computer for Gen 3 to reduce cost and increase accessibility. Enabling the system to run on an external computer means the system can be powered using existing computing resources and avoids the cost of an additional computer.

Low-cost systems are designed to increase accessibility to retinal diagnostics in areas where traditional commercial systems are not feasible. This is accomplished via tradeoffs in system features and specifications that may result in different scan protocols. The improved 40 kHz A-line rate allowed Gen 2 and Gen 3 to better replicate commercial scan protocols, such as radial and raster scan patterns, while still limiting the increase in cost. Another tradeoff that must be considered is the number of B-scans averaged to create a higher quality image. The slower A-line rate of the Gen 1 system limited the number of B-scans that could be co-registered for averaging to two. This was the motivation for restricting the HE system to N = 2 averages as well, even though it is capable of greater numbers of averaged B-scans. Gen 2 improved line rate to 40 kHz and enabled a greater number of co-registered averages to be obtained with the low-cost system. This permitted a more robust analysis of averaging with respect to CNR (see Fig. 6). For the Gen 3 system, a standard protocol of N = 9 B-scan averages was adopted. Comparing both Gen 2 and Gen 3 for N = 9 scans shows that an improvement in CNR was realized.

The self-referencing histogram matching procedure introduced for Gen 2 is a post-processing technique that improves image quality without the need for hardware upgrades. Physicians often use post-processing software for image display and analysis. While the HM algorithm has not been implemented in real time at this point, future work on optimizing and automating software for post-processing may permit real time use. Recent work by our group has enabled OCT software to be run on higher performance, less expensive computers [28], which may also enable greater capabilities for real time processing and analysis Other ways to increase access are related to scale of production. The systems presented here are research prototypes, whereas commercial systems are built in higher volumes which reduces the cost of parts. For example, if built at a higher volume, the off-the-shelf optics used here could be replaced by custom optics to improve performance and reduce the per unit cost.

5. Conclusion

OCT offers unique capabilities for non-invasive high-resolution imaging of the retina. Development of low-cost OCT is focused on improving patient access to retinal diagnostics at the point-of-care. The 3^rd generation low-cost OCT system demonstrates improved image quality as quantified by CNR, enabled by balanced detection and increases in performance of each spectrometer. Several operator usability features are also incorporated which can benefit use at the point of care. Clinical results show that low-cost OCT can produce images with quality competitive with more expensive commercial systems while being more accessible in low-resource environments.

Funding

National Institutes of Health10.13039/100000002 (R01AG072732).

Disclosures

AW: Lumedica Vision, Inc. (I,E,P), DAM: Lumedica Vision, Inc. (C). The remaining authors do not have any conflict 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. Data presented in this paper includes personal health information.