1.

Introduction

Diffuse reflectance spectroscopy (DRS) has been widely used to determine absorption and scattering properties of turbid media. When applied to biological tissue such as skin, DRS provides quantitative characterization of tissue composition for diagnostic purposes. The reduced scattering coefficient (${\mu }_s^{\prime }$) is sensitive to cells and tissue morphology, while the absorption coefficient (${\mu }_a$) relates to the biochemical content of tissue and may be used to infer the concentration of particular molecular species. In the visible and near-infrared domain, these notably include oxyhemoglobin, deoxyhemoglobin, melanin, and bilirubin, which are of great interest in a variety of clinical situations. Many groups have investigated DRS as a noninvasive tool to provide diagnostic criteria in vivo. Examples of application include tissue oxygenation monitoring,^1–3 tumor margin assessment in epithelium,^4,5 breast and colon cancer detection,^6,7 jaundice diagnosis in new-born skin,⁸ pharmacokinetics,⁹ or port-wine stain treatment evaluation.¹⁰

Spatially resolved diffuse reflectance spectroscopy (srDRS) is a particular implementation of DRS that consists of analyzing the spatial dependence of diffuse reflectance profiles produced by a given arrangement of light sources and detectors. By exploiting the spatially dependent influence of optical parameters on the diffuse reflectance intensity, determination of absorption and scattering parameters can be achieved. Furthermore, the measurement of photons at multiple source-to-detector separations (SDS) provides a potential for depth-resolved analysis.¹¹ Translation of srDRS into the clinic necessitates an optimized combination of sources and detectors, as well as modeling of light transport in tissue and specific constraints related to the targeted application.

Various ways to implement srDRS have been reported in the literature.^1,5,12–16 Most common srDRS instruments involve a broadband source and a spectrophotometer, both coupled to a fiber-optic probe ensuring light delivery and collection directly in contact with tissue. This approach has led to the development of portable instruments that have been deployed in a variety of clinical and preclinical studies.^5,16,17 Typical fiber-based probes consist of a concentric arrangement of multiple collection fibers around a central illumination fiber for maximization of signal-to-noise ratio (SNR),^5,16 as shown in Fig. 1(a). However, the fill factor and light collection efficiency of the system are limited by the density of the fibers as well as the cladding and low numerical aperture (NA) of the fibers. As a result, typically less than 10 source-to-detector separations (SDS) can be used, limiting the ability of the instrument to achieve depth specific analysis of tissue.

Fig. 1

Detection schemes for (a) fiber-based and (b) CMOS-based srDRS instruments. Views from the top (top) and side (bottom) of the instruments are provided. The diffuse reflectance at distance ${\rho }_i$ to the source is radially averaged (red-dotted lines).

Recently, alternative approaches have been investigated for further miniaturization of srDRS devices. Several groups used photodiodes (PD) in contact with tissue to collect the diffuse reflectance.^18,19 The higher NA of PDs compared to optical fibers improved light collection efficiency. However, instruments reported in these studies only involved single PD–source pairs and thus did not exploit the spatial profile of diffuse reflectance spectra. Later, this issue was addressed by Senlik and Jokerst, who developed a custom probe made of an array of multiple concentric annular PDs surrounding a central illumination aperture.¹⁴ Spatially resolved diffuse reflectance spectra were acquired at 24 SDS ranging from 400 to $3550\mu \mathrm{m}$. Such approach may have significant potential for the analysis of layered tissue using a wearable system. However, such solution involves the development of custom PD arrays, and the detection of spatially resolved diffuse reflectance is restricted to a particular geometry.

Alternatively, scattered light from tissue may be detected at a high resolution using a multipixel image sensor, typically a charge-coupled device (CCD) or a complementary metal–oxide semiconductor (CMOS) sensor. This approach was followed by several groups, who used camera-based systems to analyze the spatial dependence of diffuse reflectance spectra, either in real-space¹³ or frequency-domain.^2,3 Because of the high number of considered SDS, such systems may have significant potential to achieve depth-specific analysis of layered tissue such as skin and cervix, leading to important diagnostic criteria.^3,20,21 However, focusing optics are usually employed for remote collection of scattered photons, thereby increasing the cost and complexity of the device.

Recently, the removal of the focusing optics to achieve high-resolution imaging of scattered light from turbid samples using a contact geometry has been investigated.^22–24 Notably, Schelkanova et al.²³ employed a color CMOS sensor directly coupled to silicon phantoms and combined with a broadband source. Using this system, the authors were able to visualize subsurface microfluidic patterns mimicking epithelial vascular structures. Yet, quantitative characterization of optical parameters was not intended considering the broad bandwidth of the CMOS RGB channels.

In this work, we investigate the use of a standard high-resolution multipixel image sensor to achieve quantitative characterization of tissue optical properties in contact modality [see Fig. 1(b)]. Commercial imaging sensors based on the CMOS technology are readily available at low cost and with constantly improving optical and electronical performance. We believe that there is opportunity to exploit this technology to address a new range of clinical situations requiring cost-effective, tightly packed tissue characterization tools. In order to evaluate the potential of our approach, a prototype instrument was designed and fabricated. The instrument is made of a commercially available CMOS sensor combined with a dedicated fiber-based coupling system. The coupling system employs a surface mounted light-emitting diode (LED) and a fiber-optic plate (FOP), ensuring efficient photons transfer onto the CMOS pixel array at high spatial resolution, thereby overcoming the limitations of current fiber-based and PD-based systems. In this paper, we describe the design of the instrument, introduce the developed prototype, and finally evaluate the prototype performance on tissue-mimicking optical phantoms.

2.

Materials and Methods

2.1.

Light Sensor

In order to sample the diffuse reflectance from turbid media with high spatial resolution, we employed an eight-bit monochrome CMOS camera purchased from IDS Imaging Inc. (UI-1492-LE-M). This sensor had a $6.413\times 4.589{\mathrm{mm}}^2$ active area with $1.67\text{-}\mu \mathrm{m}$ pixel pitch. The relatively low dynamic range of detection was compensated by the use of a dynamic expansion procedure, which is detailed in Sec. 2.3.3. In the next paragraph, we describe the dedicated coupling system that was developed to produce and collect diffuse reflectance patterns in contact imaging modality. A layout of the sensor and coupling system architectures is shown in Fig. 2.

Fig. 2

Layout of the CMOS-based contact imaging system for srDRS. The imaging area (red line) is transported from the CMOS pixel area to the bottom face of the fiber optic plate, which is placed in contact with the analyzed sample.

2.3.

Experiments

2.3.1.

Optical phantoms

In order to evaluate the potential of the developed system for the characterization of absorption and scattering properties of tissue, validation experiments were conducted on homogeneous liquid phantoms, following a methodology employed by other authors.^13,15,29 Phantoms consisted of solutions of 20% Intralipid fat and black India ink (Rotring Inc., Hamburg, Germany) at varying concentrations in distilled water. The absorbance $A$ of solutions was measured by a spectrophotometer (CARY 300, Agilent) in a 1-cm tank, yielding the absorption coefficient ${\mu }_a=\ln (10)\times A$. Then, various quantities of intralipid fat were added to generate scattering in the medium. Note that the dilution of the ink due to the addition of Intralipid was taken into account in the determination of phantom absorption coefficients. For this, we applied the following dilution factor: $F=1-P/20$, where $P$ is the percentage of Intralipid in the phantom, to the absorption coefficient derived from the absorbance measurements.

Based on the work of Van Staveren et al.,³⁰ the following equations were derived to calculate the anisotropy factor $g$ and scattering coefficient ${\mu }_s$ of phantoms:

Eq. (1)

$$g(\lambda )=1.1-0.58\times {10}^{-3}\lambda ,$$

Eq. (2)

$${\mu }_s(\lambda )=2.54\times {10}^9{\lambda }^{-2.4}\times P/10,$$

where $\lambda$ is in nanometers and ${\mu }_s$ is in units of ${\mathrm{cm}}^{-1}$. The additional $P/10$ factor in Eq. (2) was used to scale the equation of Van Staveren et al. (derived for a 10%-Intralipid solution) to the percentages of Intralipid used in our study.

A set of 15 phantoms, which consisted of three groups of 1%, 2%, and 3% Intralipid concentration, respectively, referred to as IL1%, IL2%, and IL3%, was fabricated. Each group was composed of five phantoms with varying ink concentration. Corresponding reduced scattering and absorption coefficients ranged from 12.7 to $38.1{\mathrm{cm}}^{-1}$ and 0.31 to $2.30{\mathrm{cm}}^{-1}$, respectively, spanning values representative of skin properties.^31,32

2.3.2.

System characterization

The system linearity was measured by taking the average signal over the whole sensor area at constant illumination level and increasing exposure time. Background subtraction was performed for each exposure time.

For assessment of the source stability, the illumination power out of the LED was monitored over a 1.5-h acquisition period using an optical power meter (PM100D, Thorlabs).

For system noise performance characterization, repeated image acquisitions were made on the least favorable phantom, chosen as the one exhibiting the highest ${\mu }_a$ and ${\mu }_s^{\prime }$.

We calculated the SNR {SNR = 20 log[mean intensity/(standard deviation + dark current)]} out of 10 acquisitions for various exposure times. Intensity was averaged over concentric rings surrounding the illumination point, as described in Sec. 2.3.3. The dark current was measured for every exposure by taking the mean signal of dark images over the whole sensor area.

Accurate characterization of optical properties necessitates that the source radial intensity profile be accounted for. The source profile was measured using a CMOS camera (UI-1492-LE-M, IDS Imaging) and a 23-mm objective (Xenoplan 1.4/23-0902, Schneider Kreusnach). Following the procedure described in Sec. 2.3.3, images were recorded at multiple exposure times and subsequently recombined to measure the radial intensity profile produced by the LED in the plane of the FOP top face with sufficient SNR. Note that the same phantom was used for measurement of the source profile and calibration of the instrument response (see Sec. 2.4.2). The measured profile was then included in the forward model of light propagation (see Sec. 2.4.1).

Finally, flat-field inhomogeneity was corrected by using a Teflon disk with 10 mm thickness. Ten acquisitions were made and averaged. The resulting data were used as normalization factor to correct from spatial heterogeneities in FOP transmission, especially occurring near edges. In order to evaluate the inhomogeneity correction, 10 consecutive measurements of the calibration phantom were made without moving the srDRS device and the standard deviation between pixels on the same ring was compared to the measurement noise (see Sec. 3).

2.3.3.

Acquisition procedure

Consecutives steps followed during acquisitions are summarized in Fig. 3. For each measurement, reflectance images were recorded with the srDRS device using multiple exposure times. Dark images were taken with the same exposure times and subtracted from reflectance images. The resulting images were then normalized by the corresponding exposure time. This set of dark-subtracted and normalized images was employed to produce a single image, herein referred to as “high dynamic range (HDR) image,” using a custom algorithm written in Matlab (Mathworks Inc., Natick, Massachusetts). The HDR image was composed of pixels from images taken at multiple exposure times in a stepwise manner, going from the longest to the shortest exposure. At each step, pixels with values lower than the saturation limit were kept while saturated pixels were discarded and replaced by the corresponding pixels from the next image (i.e., the one recorded with the nearest lower exposure time). We set the saturation limit to 80% of the sensor dynamic so that pixel values in the HDR image remained within the linear behavior of the camera. Note that for each SDS, signals were extracted only from the image providing the best SNR. This procedure (referred to as dynamic expansion in Fig. 3) had to be employed to expand the dynamic range of detection, thereby ensuring sufficient SNR (above 20 dB) up to 4 mm from the source. Depending on phantom properties, three to five different exposure times had to be used. Especially, phantoms with highest absorption and scattering coefficients necessitated additional exposure times compared to phantoms with moderate ${\mu }_a$ and ${\mu }_s^{\prime }$. The illumination center was automatically determined through Canny filtering, following a method employed by Foschum et al.¹³ Finally, the radial symmetry of illumination was exploited to reduce noise in measured reflectance profiles. Pixel values were averaged over concentric annular areas of 10 pixels thickness around the excitation point. As rings became larger for increasing SDS, a greater number of pixels were used to calculate the reflectance and thereby the SNR drop (for a given exposure time) was slightly mitigated. This step is referred to as ring averaging in Fig. 3.

Fig. 3

Flowchart of the data processing steps for optical properties quantification. Raw reflectance images were first recorded (image acquisition). Dark images were then recorded and subtracted from reflectance images (background subtraction). This was done at multiple exposure times. The obtained images were recombined to expand the sensor dynamic (dynamic expansion). Then, pixel values were averaged over concentric annular areas surrounding the illumination center to extract the radial profile of raw diffuse reflectance (ring averaging). Theoretical reflectance profiles were derived from MC simulations and convoluted with the priory measured source beam profile. Measurement of the calibration phantom diffuse reflectance was used to yield a scaling factor through comparison with the theoretical data. The scaling factor was applied to unknown phantoms measurements to convert raw diffuse reflectance profiles into absolute unit. The obtained calibrated reflectance profiles were compared to theoretical data to extract the absorption coefficient (${\mu }_a$) and reduced scattering coefficients (${\mu }_s^{\prime }$) of phantoms.

Following the procedure described before, spatially resolved diffuse reflectance acquisitions were performed on the phantom set. Prior to measurements, phantoms were sonicated for 20 min and left at room temperature for 30 min. During acquisitions, our srDRS device was immersed in phantoms, $\sim 5\mathrm{mm}$ below the surface. Solutions were regularly stirred to prevent from surface layering of Intralipid.³³ Five measurements were made for each phantom. Between measurements, the system was removed from phantoms and gently cleaned using ethanol and water.

The acquisition was controlled using a dedicated LabVIEW interface (National Instruments Inc., Austin, Texas).

3.

Results

A prototype CMOS-based srDRS device was fabricated according to the proposed architecture (see Fig. 2). Photographs of the system are provided in Fig. 4.

Fig. 4

Photographs of the developed prototype CMOS-based contact imaging system: (a) whole system. Note that silicon treatment was performed on the printed circuit board for isolation and impermeability of the exposed electronic parts. (b) Top view of the illumination and imaging area. Dimensions are displayed in millimeters. The CMOS sensitive area delimited by the white dotted line is viewed through the fiber optic plate.

The system linearity and stability of the illumination were characterized. The maximum drift from a linear behavior was smaller than 0.6% of sensor dynamic up to 80% of saturation limit. Pixel values higher than this threshold were disregarded in our processing algorithm. The illumination power out of the LED was $240\mu \mathrm{W}$. Over the 1.5 h -acquisition period of LED stability measurement, the measured power did not drift by more than 0.2% from the nominal value.

As expected, fast decrease of SNR was observed as the radial distance to the source increased. Therefore, acquisition with multiple exposure times was necessary to maintain a sufficient SNR over several millimeters. The SNR was measured on the least favorable phantom (see Sec. 2.3.2) as a function of distance to the source. Using three to five exposure times separated by a factor close to 10 enabled to achieve SNR ranging from 20 to 35 dB up to 4 mm away from the illumination center.

Repeated measurements were used to evaluate the flat-field inhomogeneity (see Sec. 2.3.2). Depending on source–detector separation, the standard deviation between pixels on the same ring for a given measurement after inhomogeneity correction ranged from 3% to 5% of the mean signal. In comparison, the standard deviation between signals from different measurements at the same pixel ranged from 1.5% to 5% of the mean value.

Measured reflectance profiles versus best fits of modeled reflectance are shown in Fig. 5. The imaging area was divided into 211 rings, providing as many points on the reflectance profiles. For clarity, only one in five points is shown for the measured data in the radial dimension. Overall, measured data were well described by our forward model. Mean discrepancies between measured and modeled profiles calculated over all SDS were lower than 6.0%, 2.9%, and 3.2% for IL1%, IL2%, and IL3%, respectively.

Fig. 5

Measured calibrated reflectance of Intralipid phantoms (dots) are plotted against best fits of theoretical reflectance (solid lines). Measured and fitted data for a given phantom are displayed in the same color. For clarity, only one in five points is plotted for the measured data. Each window shows measured and fitted data for a group of phantoms with same Intralipid concentration and varying ink concentration.

Extracted versus expected ${\mu }_a$ and ${\mu }_s^{\prime }$ of phantoms are shown in Fig. 6. Expected properties were derived from spectrophotometer measurements and Mie theory, as described in Sec. 2.3.1, and used as reference values in error calculations. Note that we did not consider error propagation in the calculation of absolute error bars, herein corresponding to the standard deviation between properties extracted from different measurements of the same phantom. Errors were calculated both in absolute value and in percent of the expected value for the considered property and phantom. Error bars for percent errors correspond to the error bars for absolute errors expressed in percent of the expected value. To estimate the overall percent error for optical properties, we considered the percent errors (in absolute value) over the whole phantom set and calculated the average. Corresponding intervals were computed as the standard deviation between percent errors (in absolute value) for different phantoms. Optical parameters were accurately measured, with overall errors of $4.2\pm 3.5\%$ for ${\mu }_a$ and $7.0\pm 4.6\%$ for ${\mu }_s^{\prime }$. A correlation coefficient of 0.9915 was calculated between expected and measured values. Figure 6(b) shows slight underestimation of reduced scattering coefficients for highly scattering phantoms. The calculated correlation coefficient was 0.9473. For a fixed Intralipid concentration and varying ${\mu }_a$, minor variations of ${\mu }_s^{\prime }$ estimates were observed, with standard deviations of 4.9%, 5.7%, and 4.0%, respectively for IL1%, IL2%, and IL3%. We observed that errors in measured ${\mu }_s^{\prime }$ were closely related to the absorption level of phantoms. For all Intralipid concentrations, extracted ${\mu }_s^{\prime }$ decreased while ${\mu }_a$ increased. This suggests that separation of absorption and scattering influence on measured data was not complete, and some cross-talk remained between parameters.

Fig. 6

Results of phantom experiments: measured optical properties for IL1% (green), IL2% (blue), and IL3% (red) are plotted against expected values (black solid line). For each Intralipid concentration, different levels of absorption are associated to different markers (from lowest to highest ink concentration: circles, downward-pointing triangles, squares, diamonds, and upward-pointing triangles). Absolute values (top), absolute errors (middle), and relative errors (bottom) between measured and expected properties are shown for (a) absorption coefficients ${\mu }_a$ and (b) reduced scattering coefficients ${\mu }_s^{\prime }$. Error bars correspond to standard deviations between measurements.

4.

Discussion

The results obtained from phantom experiments demonstrate the ability of the developed system for accurate quantification of absorption properties of phantoms over the whole range of addressed ${\mu }_a$. However, Fig. 6(b) indicates that accuracy in ${\mu }_s^{\prime }$ extraction was correlated with the optical properties of measured phantoms. In particular, systematic underestimation of ${\mu }_s^{\prime }$ was observed for highly scattering phantom. Additionally, the impact of absorption level of phantoms on ${\mu }_s^{\prime }$ estimates was not completely removed, resulting in slight variations of extracted ${\mu }_s^{\prime }$ between weakly and highly absorbing phantoms containing the same amount of Intralipid (less than 10% overall). These errors were likely due to imperfections of our instrument calibration method. The scaling factor derived in Sec. 2.4.2 solely corrects from constant multiplicative effects of the instrument response and does not include influence of the sensor optical transfer function. Furthermore, calibration could only be efficient upon precise knowledge of the source geometry. However, immersion of the srDRS system in Intralipid solutions made accurate characterization of the source profile difficult. In our forward model, a planar interface was assumed between the medium and its environment, photons being emitted from the source at this interface. In MC simulations, this interface was placed at the limit between the FOP face and phantoms. Such description was inconsistent with the actual structure of our system. Notably, the presence of the LED dome lens, possible reflections on the FOP side, and photon trajectories between the LED chip and the planar interface were not modeled. The use of phantoms fabricated in a solid material such as poly(dimethylsiloxane)³⁷ may facilitate the measurement and modeling of the source intensity profile.

It must be noted that the contribution from dependent scattering was not considered in our study. Several authors stressed that the effect of dependent scattering should be accounted for when high concentrations of scatterers are used.^38,39 In the reported studies, neglecting this phenomenon typically resulted in higher ${\mu }_s^{\prime }$ expectations for a given Intralipid concentration, which might explain the observed underestimation of ${\mu }_s^{\prime }$ compared to expected values in our case.

Because of the limitations listed before, the developed system might show inability to address a wider range of properties (currently from 0.31 to $2.30{\mathrm{cm}}^{-1}$ in absorption and 12.7 to $38.1{\mathrm{cm}}^{-1}$ in reduced scattering). In this case, use of multiple calibration phantoms, as investigated in previous studies,^29,40 could extend the accessible ${\mu }_s^{\prime }$ range. In future work, evaluation of the device’s performance on an extended range of optical properties, notably toward lower levels of absorption, should be conducted.

A major challenge in current srDRS technology is the analysis of layered tissue with specificity in depth. In skin, epithelium and stroma are the most significant structures. The separate characterization of these layers may yield critical criteria in a variety of contexts including cancer detection,^20,21 tissue oxygenation monitoring,³ and drug permeation follow-up.⁹ Epithelium thickness varies from 100 to $300\mu \mathrm{m}$, depending on phenotype and location on the body.⁴¹ Therefore, the access to both short SDS and high resolution of detection is necessary for accurate characterization of epithelial and stromal layers. Depth-resolution is directly related to lateral resolution and therefore dependent on the apparatus used for collection and detection of photons.

We demonstrated that a multipixel sensor used in contact with tissue provided possibility to image diffuse reflectance profiles with high spatial resolution and large FOV ($\sim 20{\mathrm{mm}}^2$). Moreover, the developed srDRS coupling system provided a way to access short SDS in contact with an unmodified commercial CMOS sensor. In the current version of the system, however, presence of the protective glass lid induced blurring of the diffuse reflectance. Influence of this effect was studied through forward model calculations and only small impact was observed on obtained reflectance profiles. Differences between reflectance profiles affected or not by blurring were lower than 10% on the SDS range considered in our study. However, blurring of the diffuse reflectance may limit the ability of the system for the analysis of layered media such as skin, which require sensitivity to optical properties on reduced footprints. In this case, removal of the CMOS glass lid and positioning of the FOP directly on the pixel area may be investigated.⁴² Moreover, validation of the system on skin-mimicking multilayer phantoms should be conducted. In order to further increase the sensitivity of diffuse reflectance spectra to superficial layers, several authors have explored the use of oblique illumination.^4,43 The proposed architecture is suitable for implementation of such approach.

5.

Conclusion

A compact design for implementation of srDRS in contact with tissue has been proposed and evaluated. A prototype instrument, involving an LED source and a commercially available, unmodified CMOS sensor was fabricated. For validation of our approach, a proof-of-concept study was conducted on optical phantoms with controlled properties. Diffuse reflectance imaging was performed with high spatial resolution, large field-of-view ($\sim 20{\mathrm{mm}}^2$), and SNR comparable to that of previously developed instruments.^14,18 A dedicated coupling system involving a FOP was developed to provide access to source–detector separations as short as $480\mu \mathrm{m}$. Absorption coefficients (${\mu }_a$) and reduced scattering coefficients (${\mu }_s^{\prime }$) of phantoms were determined with 4.2% and 7.0% error, respectively, demonstrating the ability of the system for accurate quantification of tissue optical properties. Our approach is promising for the analysis of layered media such as skin and paves the way for the development of low-cost, wearable devices for skin condition diagnosis in vivo.