2.1.

Handheld Laser Speckle Imaging Device

The handheld LSI device (Fig. 1) used in all of the experiments consisted of an 8-bit, 1.32 megapixel CCD camera (CMLN-13S2M-CS, FLIR Integrated Imaging Solutions, Inc., Richmond, BC, Canada) acquiring $640\times 480\text{pixel}$ frames at 15 Hz as the imaging sensor. An imaging field of view (FOV) of $\sim 90\mathrm{mm}\times 67.5\mathrm{mm}$ ($4:3$ ratio) at an imaging distance of 300 mm was obtained using a C-mount lens (Computar C-Mount 13-130 mm Varifocal Lens, Computar). Imaging parameters were selected to achieve $\sim 3\text{pixels}$ per speckle, thus satisfying the Nyquist sampling criterion.¹⁰ The data were collected and processed on a tablet computer (Microsoft Surface Pro 2, Microsoft Inc.) using a custom-written graphical user interface (GUI) in MATLAB^® (The Mathworks, Natick, Massachusetts).

Fig. 1

Assembled handheld LSI device. (a) Fully assembled device and (b) device in a tripod-mounted setup.

The coherent light source was an 809 nm near-infrared laser diode (140 mW, Ondax Inc., Monrovia, California) positioned at a slightly oblique angle to assist with alignment. A 300-mm ruler was used to check the distance from the LSI device and the sample being measured. If the device was too close during data acquisition, the irradiated region would shift toward the upper left quadrant in the FOV; if the device was too far, the irradiated region would shift toward the lower left quadrant. Data acquisition was initiated only after the irradiation region was in the approximate center of the FOV. The exposure time used for each image was 5 ms, based on the findings of Farraro et al.⁸

The tablet was placed inside a protective case (Urban Armor Gear), which was modified to allow direct mounting of the lens, camera, and laser to the tablet. The device was designed for use in both mounted and handheld configurations.

2.2.

Fiducial Marker

The FM was an 18% reflectance grey card (Neewer, model #10079934) commonly used in photography. The FM was used to sort and align the acquired images (see Sec. 2.3). It was attached to the surface of a solid silicone phantom (see Sec. 2.4) and imaged using our LSI device in mounted and handheld setups [Fig. 2(a)]. Each data set contained 150 images for analysis. All images were converted into spatial speckle contrast ($K$) images using a $7\times 7\text{pixel}$ sliding window with the equation $K=\sigma /⟨I⟩$, where $K$ is the contrast, $\sigma$ is the standard deviation within the window, and $⟨I⟩$ is the mean intensity of the pixels contained within the window.^1,2,4,6 $K_{\mathrm{FM}}$ was quantified from each image.

Fig. 2

FM included into imaging protocol allows for motion artifact detection based on speckle contrast ($K$). (a) The 18% gray card used in our imaging protocol was placed in the lower left corner of all frames during data acquisition to allow for sorting based on motion artifact and image alignment when calculating average speckle contrast images. Here, the FM was placed on the surface of a flow phantom. The FM was identified and the speckle contrast value of the FM ($K_{\mathrm{FM}}$) was quantified and plotted in descending order for all images within each data set. (b) The handheld $K_{\mathrm{FM}}$ values plotted show a rapid decline within a representative sorted handheld data set. The $K_{\mathrm{FM}}$ mounted values plotted remain relatively stable across all images within a sorted mounted data set. (c) The table shows candidate values of $K_{\mathrm{FM},\mathrm{thres}}$, the percent difference between each of these values and the known $K_{\mathrm{FM}}$ value, and the number of images in the handheld data set that exceeds each of the $K_{\mathrm{FM},\mathrm{thres}}$ values.

2.3.

Estimation of Motion Artifact and Image Alignment Using FM

Motion artifact was estimated using the quantified $K_{\mathrm{FM}}$ in each image. All images within a data set were sorted in descending order based on the $K_{\mathrm{FM}}$ associated with each image. This approach was based on the assumption that with increasing motion artifact, $K_{\mathrm{FM}}$ would decrease from the “true” value (0.50, in this study) measured with the mounted LSI setup. We set arbitrary threshold $K_{\mathrm{FM}}$ differences of 10% and 20% (i.e., $K_{\mathrm{FM},\mathrm{thres}}=0.45$ and 0.40, respectively) and determined the number of images in the collected data sets whose $K_{\mathrm{FM}}$ values exceeded the threshold, as well as the corresponding $K$ values of a specified ROI extracted from the subset of images (see Secs. 2.4 and 2.5). As a comparison, average $K$ images were also created using all 150 images to compare differences in average $K_{\mathrm{FM}}$ values using the FM-based approach and an unsupervised approach.

Prior to ROI extraction from a given subset of LSI images, we used the FM to align the LSI images using custom-written MATLAB^® software. First, a median sliding filter with a $7\times 7\text{pixel}$ window filter was run on each of the images. Next, the software calculated the necessary transformation matrices. The transformation matrices were found by performing a translation, rotation, and scaling in multiple axes to align each raw image to the highest ranked raw image.¹¹ These transformation matrices were created by combining MATLAB^® functions with our custom-written software. We converted the raw images to spatial $K$ images and then used the transformation matrices on their respective $K$ images.

2.4.

In Vitro Flow Phantom

Flow phantom experiments were performed using a solid silicone phantom created with the methods outlined in Ayers et al.¹² We also incorporated a clear plastic tube (diameter 10 mm) as a surface-level inclusion flow tube. The flow medium was a 1% Intralipid solution (Fresenius Kabi) infused into the flow tube using a mechanical pump (NE-1000 Single Syringe Pump, Pump Systems Inc.). Two ranges of flow speeds were used: (1) 0.0 to $1.0\mathrm{mm}/\mathrm{s}$, in $0.2\mathrm{mm}/\mathrm{s}$ increments and (2) 1 to $5\mathrm{mm}/\mathrm{s}$ in $1\mathrm{mm}/\mathrm{s}$ increments. Sequences of 150 images were acquired using mounted and handheld setups at each flow speed. As described in the previous section, $K_{\mathrm{FM}}$ was identified based on specified threshold values and aligned using the FM, followed by quantification of $K$ within an ROI selected inside the flow tube ($K_{\mathrm{FLOW}}$). The difference in $K_{\mathrm{FLOW}}$ resulting from data collected with mounted and handheld setups was determined. Bland–Altman analysis¹³ was performed with Prism 7 software (GraphPad Software, Inc.) to study the measurement performance of the handheld and mounted LSI setups.

2.5.

In Vivo Mounted Versus Handheld Imaging of Porcine Burn Wound Model

All experiments were performed in accordance with the Animal Care Use Committee of University of California, Irvine (IACUC # 2015-3154). We performed a graded burn wound experiment using a porcine model described previously.⁵ Briefly, 30-mm-diameter burns were created using a brass tool heated to 100°C. The burn severity was based on the contact time of the burn tool to the porcine skin. Contact time was varied from 5 s (superficial) to 40 s (full thickness) to generate a full range of burn severities (16 burns per pig, two pigs). For in vivo validation of the handheld LSI approach, we collected data from both superficial and full thickness burns. Each data set was acquired with mounted and handheld LSI setups and contained 100 images to reduce acquisition time. The FM was placed in the FOV of the images and $K_{\mathrm{FM}}$ calculated from each image. An average $K$ image was created using the sort, threshold, and align methods described above. Quantification of $K$ from an ROI within each burn ($K_{\mathrm{BURN}}$) was performed and compared for mounted and handheld setups.

3.1.

$K_{\mathrm{FM}}$ is Indicative of Motion Artifact

To investigate the influence of motion artifact on handheld LSI measurements, the FM was included in all images acquired for each data set [Fig. 2(a)]. The $K_{\mathrm{FM}}$ on a static phantom was quantified with mounted and handheld setups. The true $K_{\mathrm{FM}}$ measured with the mounted LSI setup was 0.50. After imaging the marker with these solid phantoms, we sorted the $K$ values of the FM in descending order. $K_{\mathrm{FM}}$ decreased more quickly in handheld measurements than in mounted measurements [Fig. 2(b)]. The relative stability of the $K_{\mathrm{FM}}$ across the mounted data set [Fig. 2(b)], compared with a handheld data set [Fig. 2(b)], demonstrated the sensitivity of $K$ values to motion artifact. The reduction in $K_{\mathrm{FM}}$ in the mounted data set was associated with the button press on the tablet screen to initiate data acquisition. The image with the highest $K_{\mathrm{FM}}$ value was considered the “best” image and contained the least motion artifact ($K_{\text{best}}$). The image with the lowest $K_{\mathrm{FM}}$ value was considered the “worst” image and contained the most motion artifact ($K_{\text{worst}}$). To analyze the range of motion artifact within each data set, the percent difference between $K_{\text{best}}$ and $K_{\text{worst}}$ was calculated. In the mounted setup, the percent difference was found to be 8%, whereas the percent difference in the handheld setup was 52%. Furthermore, $K_{\text{best}}$ measured with the handheld setup (0.47) was lower than the true $K_{\mathrm{FM}}$ value, demonstrating that, even in the best-case comparison, user motion introduces an error.

We then applied threshold $K_{\mathrm{FM}}$ ($K_{\mathrm{FM},\mathrm{thres}}$) values to identify subsets of images with $K_{\mathrm{FM}}$ differences ranging between 0% and 20% [Fig. 2(c)]. As expected, the number of images in each subset decreased as $K_{\mathrm{FM},\mathrm{thres}}$ approached the true $K_{\mathrm{FM}}$ value. The maximum $K_{\mathrm{FM}}$ value from the handheld data sets was consistently below the true $K_{\mathrm{FM}}$ value, demonstrating that some motion artifact was present within each image of the handheld data sets. Collectively, these data demonstrate the issues related to motion artifact. If we were to use all of the images within a handheld data set to quantify the mean $K$ of an ROI, the handheld setup would potentially result in grossly inaccurate $K$ values.

3.2.

Using the FM to Account for Motion Artifact Improves Accuracy of K Values Calculated from Data Acquired with the Handheld LSI Setup

To investigate how the FM can help with motion artifact correction, the image processing protocol outlined in Sec. 2.3 was applied to use the FM to identify images with the least motion artifact. We first compared the mounted LSI data with $K_{\mathrm{FM}}=0.50$ (i.e., data with no motion artifact) with handheld LSI data using all images, to mimic unsupervised analysis of the data; the median difference in $K$ values from an ROI centered on the flow tube inclusion ($K_{\mathrm{FLOW}}$) was 20% (range: 7% to 26%) [Fig. 3(a)]. We then applied $K_{\mathrm{FM},\mathrm{thres}}$ values of 0.45 and 0.40 to extract subsets of images with $K_{\mathrm{FM}}$ values above these thresholds to analyze further. Each image subset was then aligned using the FM. Using $K_{\mathrm{FM},\mathrm{thres}}$ of 0.45 and 0.40 resulted in median differences in $K_{\mathrm{FLOW}}$ values of 5% (range: 1% to 12%) and 8% (range: 0.3% to 16%), respectively [Fig. 3(b)], demonstrating the improved accuracy resulting from use of the FM. Bland–Altman analysis with the two thresholds showed biases in $K_{\mathrm{FLOW}}$ values of $-0.0089$ and $-0.014$ for $K_{\mathrm{FM},\mathrm{thres}}$ of 0.45 and 0.40, respectively [Figs. 3(c) and 3(d)]. Since the 95% confidence limits of agreement with both $K_{\mathrm{FM},\mathrm{thres}}$ values are similar ($-0.023$ to 0.0047 for $K_{\mathrm{FM},\mathrm{thres}}=0.45$; $-0.031$ to 0.0029 for $K_{\mathrm{FM},\mathrm{thres}}=0.40$), we selected $K_{\mathrm{FM},\mathrm{thres}}=0.40$ to apply in the subsequent in vivo study described in the next section.

Fig. 3

$K$ values of handheld and mounted setups from in vitro flow phantom experiments. (a) With an unsupervised approach, use of all 150 handheld LSI images resulted in an error of up to 20%. (b) When using the FM to select a subset of images with a minimum $K_{\mathrm{FM}}$ value to account for motion artifact and to realign images, the accuracy of handheld LSI improves, with errors of 8% and 5% for $K_{\mathrm{FM}}$ values of 0.40 and 0.45, respectively. (c) Bland–Altman plot of $K_{\mathrm{FM}}$ data collected using $K_{\mathrm{FM},\mathrm{thres}}=0.45$. We observed a systematic bias of $-0.0089$ between the two measurement approaches (95% confidence limits of $\text{agreement}=-0.023$ to 0.0047. (d) Bland–Altman plot of $K_{\mathrm{FM}}$ data collected using $K_{\mathrm{FM},\mathrm{thres}}=0.40$. We observed a systematic bias of $-0.014$ between the two measurement approaches (95% confidence limits of $\text{agreement}=-0.031$ to 0.0029).

3.3.

Accounting for Motion Artifact with the FM Improved Handheld LSI Performance: In Vivo Porcine Burn Model

To determine the translational potential of the in-vitro findings to in-vivo use, we imaged porcine burn wounds. Figures 4(a) and 4(b) show a comparison of speckle contrast maps calculated from speckle image sequences captured with the mounted LSI setup, either without [Fig. 4(a)] or with [Fig. 4(b)] the use of the FM to align the images. Without alignment, the speckle contrast map appears blurred due to movement occurring during data collection. In this specific case, the $K$ values were only slightly affected without the use of the FM [median difference in $K$ values of 2% (range: 0.1% to 4.0%)], which may have been due to the homogeneity of $K$ values within the burned region. In a more heterogeneous flow map, the accuracy is expected to degrade without the use of an alignment technique.

Fig. 4

Mounted versus handheld speckle contrast of superficial and full thickness burns induced on a porcine model. For both the mounted and handheld data sets, we applied $K_{\mathrm{FM},\mathrm{thres}}=0.40$. (a, b) Average $K$ image of sequence of mounted LSI images of a full-thickness burn wound, (a) without and (b) with the use of the FM for alignment and thresholding. Within the image, the larger circular region is the burn region, and the smaller four circular regions are biopsy points taken over the course of the study. (c) Day 1 $K_{\mathrm{BURN}}$ data were obtained at $\sim 24\mathrm{h}$ postburn. The difference in $K_{\mathrm{BURN}}$ between thresholded mounted and handheld data sets was 4%, and the difference between superficial and full thickness burns was 27%. (d) Day 4 $K_{\mathrm{BURN}}$ data were acquired 4 days postburn, when the burn wounds have stabilized. The difference in $K_{\mathrm{BURN}}$ between thresholded mounted and handheld data sets was 1%, and the difference between superficial and full thickness burns was 32%. The uncorrected handheld data reported $K_{\mathrm{BURN}}$ values $\sim 20\%$ lower than those reported from handheld data corrected with the use of the FM.

We used our FM and processed the handheld data using $K_{\mathrm{FM},\mathrm{thres}}$ of 0.40. This resulted in $10.2\pm 5.4$ images used to create the speckle contrast map. Due to motion artifact resulting from subject breathing, $K_{\text{best}}$ collected with the mounted setup was 0.46, which is less than the true FM value of 0.50. Hence, we also applied $K_{\mathrm{FM},\mathrm{thres}}$ of 0.40 to the mounted LSI data sets.

By imaging superficial and full thickness burns using the LSI device in mounted and handheld setups, we determined the magnitude of $K$ values measured within the burn ($K_{\mathrm{BURN}}$) for both imaging setups. On day 1 postburn, the percent difference of $K_{\mathrm{BURN}}$ between the mounted and handheld data was 4% for each burn type. In contrast, $K_{\mathrm{BURN}}$ of full-thickness burns was 27% higher than that of superficial burns [Fig. 4(c)].

On the final day (day 4) of imaging, the burn wounds have stabilized and are more indicative of the burn type.⁵ The percent difference of $K_{\mathrm{BURN}}$ on day 4 for mounted versus handheld was $\sim 1\%$. $K_{\mathrm{BURN}}$ of full-thickness burns was 32% higher than that of superficial burns. These data suggest that, regardless of mounted or handheld LSI setup, we still were able to use $K$ values alone to differentiate between a superficial and full thickness burn. Without the use of the FM to threshold and align the LSI images, handheld LSI reported $K_{\mathrm{BURN}}$ values that were $\sim 20\%$ lower (range: 16% to 25%) than handheld LSI with the use of the FM.