Component and system evaluation for the development of a handheld point-of-care spatial frequency domain imaging (SFDI) device

Recently, digital photography has become an efficient and economic method to assist dermatologists in monitoring skin characteristics. Although this technology has advanced a great deal in resolution and costs, conventional digital cameras continue to only provide qualitative recording of color information. To address this issue, we are developing a compact, quantitative skin imaging camera by employing spatial frequency domain imaging (SFDI), a non-contact approach for determining tissue optical properties over a wide field-of-view. SFDI uses knowledge of optical properties at multiple wavelengths to recover concentrations of tissue constituents such as oxy/deoxy-hemoglobin, water, and melanin. This method has been well researched and presented in laboratory and research settings. The next step in the development of SFDI systems is to make typical systems compact and cheaper using commercial components. We present our findings by performing a component-by-component analysis of key SFDI system components including light sources, projectors, and cameras.


INTRODUCTION
The most prominent technique to monitor tissue health in medicine remains visual inspection. The color and feel of a tissue can tell clinicians a great deal about the health of a tissue. In the same spirit, digital cameras have also become a tool to aid clinicians in monitoring various skin conditions. Cameras have become an increasingly useful tool to record tissue appearance, provide surgical guidance, and aid in assessment. Digital cameras are relatively low cost, based on advances in consumer electronics, and compact, which makes them easy to transport and store within a clinical environment.
Although digital camera technology has advanced tremendously and is pervasive in the consumer market, the data obtained is generally qualitative due to dependence on lighting conditions. Even in identical lighting conditions, a color image recorded by two similar cameras for the same object may vary significantly [1] . For certain types of skin phenomena, this camera-dependent effect on color data may change the interpretation of clinical images [2] . Varying lighting conditions can affect the appearance of an object on the same camera as well. This suggests that quantitative, objective imaging tools could have tremendous impact in clinical applications such as dermatology, which has already demonstrated a willingness to use technologies such as dermoscopy, to improve clinical assessment of suspicious tissue [3] .
In recent years, a non-contact quantitative imaging technique known as spatial frequency domain imaging (SFDI) has been invented and developed at the Beckman Laser Institute. SFDI is a non-contact, wide-field imaging modality that can measure the absolute concentration of chromophores in tissue. SFDI is able to make quantitative measurements by using a well-calibrated multi-spectral light source coupled with a patterned illumination scheme and camera-based Chromophores that are typically measured in tissue include oxy/deoxy-hemoglobin, melanin, and water. SFDI has been applied to a number of skin applications including the assessment of port wine stain response to laser therapy [4] , characterizing tattoo optical properties [5] , assessing burn severity [6] , and monitoring the health of skin flaps [7] . Figure 1 shows a high-level illustration with the key components needed to build a basic SFDI system. These components include a projector, light source, and camera.
An extensive description and analysis of SFDI has been described by Cuccia et al [8] . In short, light passes through a spatial light modulator (SLM) to generate spatially modulated sinusoidal patterns, which are then projected onto the tissue. In many cases, the SLM employed is a digital micromirror device (DMD), which is commonly used in commercial projectors [8,9] . The diffusely reflected light from the tissue enters the camera lens, and is detected by the charge-coupled device (CCD). Multiple patterns of different spatial frequencies are required to recover tissue optical properties. In order to recover the AC amplitude, we use a demodulation method first described by Neil et al [10] . In this embodiment, sinusoidal illumination patterns of a certain frequency are sequentially projected onto tissue with three phase shifts (0, 120, 240 degrees), and the amplitude envelope for a given spatial frequency is recovered using a demodulation term. The multiple pattern requirement makes the SLM a key component of a SFDI system. This remitted light data is calibrated to a reference phantom with known optical properties to recover a calibrated reflectance. From the calibrated diffuse reflectance data at multiple frequencies, we can obtain the reduced scattering and absorption coefficients using Monte Carlo [9] or diffusion light transport models in the spatial frequency domain. We can then derive relevant chromophore concentrations based on the derived absorption coefficients at specific wavelengths using Beer's Law [11] . The spectral and intensity stability of the light source and camera are crucial in this calibration step as fluctuations between calibration and measurement can be misinterpreted as incorrect optical properties of the tissue sample.
Our lab has developed many generations of instrumentation. Typically, all components in our systems have been scientific grade and or in the form of developer kits with minimal focus for size and cost. Our group has also typically focused on near-infrared (NIR) illumination as it typically interrogates tissue deeper than the naked eye due to the relatively low absorption of blood and water. We feel that deeper detection is advantageous for early disease detection and we require that NIR illumination is a design requirement for new systems. Here, we propose a component evaluation as initial steps towards the development of a compact SFDI device in a similar form factor as a handheld digital camera. Our goal is to integrate the functionality of previous SFDI systems we have developed into an inexpensive, compact system tailored to clinical use for dermatology in particular.
When evaluating components for developing a handheld SFDI system, there are several considerations that are needed for each component. In this paper, we will focus primarily on grayscale linearity, light source stability, and spectral throughput. As described above, SFDI derives chromophore concentration estimates based on data taken from multiple spatial frequencies using a demodulation method. This demodulation equation assumes true sinusoidal patterns and is a requirement for any SLM. In our work with commercial projectors, we have found that this necessitates a calibration step by measuring the linear grayscale intensity output versus input curve. Also, SFDI requires data taken at specific known and stable wavelengths. Therefore, the light sources that we employ must have good spectral stability over time so that our measurements are accurate. Also, a key practical issue for any clinical measurement is artifacts due to patient motion. Although our group has developed algorithms to account for this at some level, it is still ideal to minimize our imaging times [12] . Our approach has focused on using discrete, bright sources such as light emitting diodes  (LEDS), and penetration d considered. duration of th

Projecto
As described frequencies.
As described using a DMD throughput.
Our lab ha typically exp potential opti such as the L

Light So
In order t correspondin One way of a wavelengths. typically requ second option option for a l susceptible to more desirab patterns, and module.  In order to OneLight sou wavelengths onto the CCD nm?). This w output. Both accounted fo

Projecto
In general grayscale out grayscale out projectors tha given project 6(a)). ectrum for eac ings, we also in did this by dri rated in Section a tion arm of SF ng imaging time grade camera cameras, as we wn in Figure 4 reuznach Cineg aimed a broadb  After applying grayscale correction (described in the previous section) to the LightCrafter, the grayscale intensity curve became linear. As a result, the demodulated images we obtained were far better in quality when compared to those obtained without grayscale correction, as will be seen in Figure 6. The non-linearity in the initial grayscale output versus input curve is reflected in the detected sinusoidal pattern. It should be noted that this data was taken from a 99% Spectralon reflectance standard at a spatial frequency of 0.05 mm -1 , which is a common spatial frequency that we use to make SFDI measurements.

Ratio of Output to Input
Spectral throughput data for the LightCrafter light engine is shown in Figure 7. Data is shown only for the LightCrafter because it is suitable for a handheld system, and is therefore applicable to this study. The ratio of output to input as a function of wavelength is given in Figure 7, which shows a clear degradation in output intensity beginning at around 700 nm, and decaying rapidly up to 900 nm, where there is virtually no signal. The ramifications of this degradation in the NIR region are significant, and will be discussed in Section 4.

Light Sources
The temporal spectral stability data acquired from the LED modules is shown in Figure 8. For the LightCrafter LEDs, we measured peak wavelengths at 470, 535, and 625 nm, which correspond to the peak wavelengths of the blue, green, and red LEDs respectively. For the Roithner 970 nm module, we measured a peak wavelength at 970 nm, as expected. Over the course of an hour, we observed an overall reduction in measured intensity of about 10% for the visible LEDs, and 15% for the NIR LED. Over that same period of time, we see minimal shifting of the peak wavelength for the visible LEDs, and a red-shifting of approximately 10 nm for the NIR LED. These changes occurred predominantly within the first 15 minutes after powering up. The decrease in amplitude and wavelength shifting were a consequence of the heating of the LEDs. One way to account for these effects in an SFDI system is to use thermo-electric cooling (TEC) [13] . However, implementation of TEC in a handheld system is not ideal, as it can add to the form factor of the final device. We will discuss possible alternatives to TEC in Section 4. function of input current. As input currents were increased in increments of one ampere, the amount of heat generated by the component was greater, so the shifting was also likely due to heating.

Camera
Similar to the projector throughput experiment, we also analyzed the throughput of two camera lenses (Schneider Kreuznach and Panasonic Lumix). This data is depicted in Figure 10a. We show that the throughput of the researchgrade (Kreuznach) lens is excellent for all wavelengths tested, and that the consumer-grade (Lumix) lens throughput degrades in the NIR region, presumably a consequence of a manufacturer coating.
In Figure 10b, we show a plot of CCD throughput versus wavelength. The metric that we use to quantify throughput is relative efficiency, which is related to the number of photons converted to electrons, which are used to determine intensity. Figure 10b shows a significant degradation in relative efficiency in the NIR region, resulting in a greater than five-fold decrease from 500 nm to 900 nm. We also investigated the effect of increasing current to the LED. These results are shown in Figure 9. We observed that as input current is increased, the output intensity increased. However, the spectrum was also red-shifted as a In Section 3.1, we presented results pertaining to the grayscale intensity curves of COTS projectors. The data obtained using SFDI relies heavily on the precise spatial modulation of light. That is, the sinusoidal intensity patterns applied to the sample must be actual sinusoids. Therefore, it is critical that the SLM in an SFDI device produces accurate sinusoidal patterns. In COTS projection units, the grayscale output curve is often non-linear in order to account for the non-linear visual response curve of the human eye. This non-linearity makes sense for the purpose of video projection, but results in SFDI data with distinct artifacts. The grayscale calibration method proposed has shown to be a viable option to overcome this effect.
Another area of emphasis in this study has been the requirement to interrogate samples at specific wavelengths. This is directly related to the results obtained in Section 3.2, which highlights the spectral stability of visible and NIR LED modules. We found here that attenuation and spectral shifting over time were prevalent. We also showed that spectral shifting was present as we increased input current. In each case, we hypothesize that this shifting is due to heating of the LEDs, to which TEC was not applied due to size constraints. A barrier to developing a handheld device will be addressing these thermal issues. In order to account for temporal shifting, one possible solution may be to integrate a "standby" modality in order to allow the LEDs to warm up. We found that most of the shifting occurred during the first 15 minutes after power up, so another solution could be to simply wait for 15 minutes before using the device. This may be cumbersome, however, as a point-of-care device should have properties that allow for on-the-fly availability.
In Sections 3.1 and 3.3, we showed the limited spectral throughput for various components. Spectral throughput of components in the light path is directly related to the integrity of the obtained reflectance data and is ultimately used to compute chromophore concentrations. This will become even more relevant when cross polarization, which is used to suppress specular reflection from the sample surface, is integrated into the system. For rough surfaces such as skin, cross polarization is necessary for subsurface interrogation [8] . Since throughput in the NIR region is essential to calculating the concentration of many chromophores (i.e. oxy/deoxy-hemoglobin, water, etc.), we will eventually need to address the lack of NIR throughput issue prevalent in projectors, lenses, and detectors. One option may be to simply find other components whose internal optics do not reject NIR signals. Another option could be to remove the anti-IR coating on the internal optics, or replace these internal components with uncoated components.
In order to evaluate the performance of our device, we must run tests on phantoms with known optical properties that closely mimic human skin. We have already devised a method for fabricating such phantoms [14] . In future studies, these phantoms will be used to evaluate the performance of our system in the context of a clinical scenario.

CONCLUSION
We have analyzed several compact and low cost hardware components, and have presented data relevant to the component evaluation for implementation of a handheld point-of-care SFDI device. We have investigated system characteristics including grayscale linearity, spectral stability, and throughput, and related them to how they affect SFDI meaasurements. Although we will use this information to develop a specific device, one can apply the same experimental framework to develop most if not all SFDI devices.