Development of an imaging device for label-free parathyroid gland identification and vascularity assessment

Abstract

During thyroid surgeries, it is important for surgeons to accurately identify healthy parathyroid glands and assess their vascularity to preserve their function postoperatively, thus preventing hypoparathyroidism and hypocalcemia. Near infrared autofluorescence detection enables parathyroid identification, while laser speckle contrast imaging allows assessment of parathyroid vascularity. Here, we present an imaging system combining the two techniques to perform both functions, simultaneously and label-free. An algorithm to automate the segmentation of a parathyroid gland in the fluorescence image to determine its average speckle contrast is also presented, reducing a barrier to clinical translation. Results from imaging ex vivo tissue samples show that the algorithm is equivalent to manual segmentation. Intraoperative images from representative procedures are presented showing successful implementation of the device to identify and assess vascularity of healthy and diseased parathyroid glands.

Graphical Abstract

A new imaging device is presented for the purpose of identifying parathyroid glands and assessing their vascularity during surgery, without the need for contrast agents, through the combination of autofluorescence imaging with laser speckle contrast imaging. The device employs an algorithm to analyze images obtained and automate the delivery of relevant information to the surgeon, reducing a barrier to clinical translation.

1. INTRODUCTION

Hypocalcemia is a common occurrence after total thyroidectomy. Reported rates vary partly due to differences in definitions [1–3], with some as high as 47% [4], however one review places the median reported rate at 27% [5]. Post-thyroidectomy hypocalcemia occurs when healthy parathyroid glands, which are the human body’s main calcium regulators [6], are accidentally excised, injured or have their blood supply disrupted during thyroid resection. This leads to a loss of parathyroid hormone (hypoparathyroidism), which is responsible for signaling the body to increase calcium levels in the bloodstream when they are low. Calcium is required for a host of functions including muscle contraction and neuronal excitability. Hence, patients with prolonged low levels will suffer from tetany, cardiac arrhythmias and muscle spasms [6]. These patients therefore require calcium supplementation and repeated hospital visits in order to maintain normal calcium levels, and in about 5% of total thyroidectomy patients the condition is lifelong [2], making this a huge economic burden for these patients [7].

The two main challenges in the preservation of healthy parathyroid glands during thyroidectomy are the ability to identify the glands and assess their vascularity. The parathyroid glands are small (on the order of 3–8 mm [8]) and appear visually similar to adjacent tissues in the neck, making identification challenging. Paras et al. were the first to report autofluorescence of parathyroid glands under near-infrared illumination [9]. Using 785 nm excitation, they obtained fluorescence spectra peaking between 820–830 nm that were at least twice as intense as the spectra from the thyroid. Further work showed that the ratio of the parathyroid fluorescence to that of the thyroid ranges from 1.2 to above 20 [10, 11]. These studies relied on point measurements of fluorescence spectra using a fiber-optic probe and led to the development of a clinical device for probe-based parathyroid detection [12]. Another approach to detecting the parathyroid near-infrared autofluorescence (NIRAF) is based on imaging, typically using a longpass filter in front of a camera in order to visualize the fluorescence [13–17]. On one hand, the probe-based approach provides spectral information and lends itself more readily to quantifying fluorescence intensity. Measurements are made by being in contact with the tissue of interest, and therefore it is more likely to detect weaker fluorescence signals. It is also quite similar in handling to nerve-monitoring devices used routinely in thyroid surgeries, making it easier for surgeons to adopt. On the other hand, the imaging approach provides spatial information which can be very useful when surveying the surgical field. There is also evidence to suggest that the imaging approach can detect parathyroid glands obscured by fat or connective tissue [18]. Both approaches have demonstrated the consistency of parathyroid NIRAF and its utility in helping prevent accidental parathyroidectomy.

The second challenge of assessing parathyroid gland vascularity is just as important for positive postoperative outcomes. If a parathyroid gland is devascularized, that is, its blood supply is completely severed during the surgery, it will not be able to produce and secrete parathyroid hormone postoperatively. A devascularized parathyroid gland can however be salvaged intraoperatively by autotransplantation, a process whereby the gland is cut into tiny pieces and inserted into a muscle pouch [19]. Accurate knowledge of parathyroid vascularity is therefore needed to aid decision-making regarding autotransplantation. Visual assessment and perturbation of the parathyroid to elicit bleeding are the most commonly used methods to assess parathyroid vascularity. Indocyanine green (ICG) angiography is a method that is currently gaining acceptance in this field. It has been reported that observing one parathyroid gland with strong ICG fluorescence is sufficient for the patient to have normal parathyroid hormone levels post-thyroidectomy [20]. However, since this method relies on the fluorescence of ICG which is in a similar spectral region and much stronger than the parathyroid autofluorescence, autofluorescence-based parathyroid detection cannot be accomplished post ICG injection. Current use of ICG to assess parathyroid viability also faces a few other challenges. With a half-life of 3.4 minutes, repeated injections of ICG may sometimes be required depending on the duration of the surgery [20]. Another problem is that the current method of evaluation relies on a qualitative scoring of fluorescence intensity by the surgeon, which makes it difficult to standardize measurements across surgeons [20]. Finally, a small percentage of patients can suffer allergic reactions to the dye [21]. There is therefore still a need for a label-free objective approach to assess parathyroid vascularity.

An alternative method to assess parathyroid vascularity without the use of contrast agents is laser speckle contrast imaging (LSCI). This technique analyzes the interference pattern produced when coherent light is incident on a surface. Motion causes blurring of this interference or speckle pattern, which can be analyzed spatially or temporally to produce, for instance, blood flow maps in biological tissue. To analyze the speckle pattern, speckle contrast, which represents the ratio of the standard deviation of pixel intensities within local regions to their corresponding mean intensities, is calculated. For temporal analysis, standard deviations and means are calculated for each pixel over a number of images, while for spatial analysis, they are calculated per image within a sliding window (typically 5-by-5 or 7-by-7 pixels) [22]. Lower speckle contrast is indicative of greater blurring of the speckle pattern and therefore greater blood flow. We have recently demonstrated that LSCI is able to accurately distinguish between parathyroid glands classified as well-vascularized versus vascularly damaged, as compared to the eye of an experienced surgeon. Well-vascularized glands consistently had lower speckle contrast [23].

The purpose of this work is to demonstrate a device – the parathyroid speckle and autofluorescence imager (ParaSPAI) – that combines NIRAF imaging with LSCI to enable simultaneous label-free parathyroid identification and vascularity assessment. An added benefit of this combination is the ability to use fluorescence images to automatically segment parathyroid glands and determine their speckle contrast. Previous work required manually segmenting speckle contrast images in order to obtain this information [23], which could be a barrier to clinical translation. Here, we present a combined autofluorescence and LSCI system, as well as a segmentation algorithm developed to automate the process of determining parathyroid gland speckle contrast. The device and algorithm were evaluated first on ex vivo tissue samples, and then intraoperatively in 3 patients.

2. METHODS

2.1. Device design

The ParaSPAI was developed in-house and is depicted in Figure 1. The device was constructed on a cart that can be wheeled into and out of the operating room (Figure 1.A). It has a single mode 785 nm diode laser with 60 mW power output (Innovative Photonics Solutions, Monmouth Junction, NJ) as its illumination source. An articulated arm (ICWUSA, Medford, OR) is attached to the cart, capable of extending about 4 feet from its edge, with an attachment for a sterile handle to allow maneuvering by the surgeon. The imaging head that acquires both LSCI and autofluorescence images is attached to the end of this articulated arm. A single mode fiber optic patch cable (Thorlabs, Newton, NJ) couples light from the laser to a lens tube attached on the exterior of the imaging head, illuminating a ~30 mm diameter spot at a distance of 400 mm from the imaging head. A linear polarizer (Thorlabs, Newton, NJ) in the lens tube is used to polarize its output to enable reduction of specular reflection in acquired images. The illumination has an approximately Gaussian profile and the maximum power density across the spot was measured to be 6.8 mW/cm². Two laser pointers, one green and one red (Digi-Key, Thief River Falls, MN), attached on the sides of the imaging head guide the surgeon in positioning the device so that the tissue of interest in roughly in the center of the field of view and in focus when imaging. The laser pointers are angled such that their beams overlap at a distance of 400 mm from the imaging head.

FIGURE 1.

Schematic of the ParaSPAI (A); schematic of imaging head (B); optical layout of system (C). DBS = dichroic beamsplitter, L1 = lens 1, L2 = lens 2, LP = linear polarizer, LPF = longpass filter, M = mirror, NDF = neutral density filter

Light scattered from the tissue is detected by the imaging head through the LSCI or NIRAF path (Figure 1.C). Both paths share a collection lens (Edmund Optics, Barrington, NJ) which collimates light from the imaging plane onto a ~800 nm dichroic beamsplitter (Semrock, Rochester, NY). In the LSCI path, the laser light scattered from the tissue reflects off the dichroic and is focused onto a near-infrared-optimized CMOS camera (acA1300–60gmNIR, Basler AG, Ahrensburg, Germany), operated at 8-bit depth. After the dichroic mirror, there is a linear polarizer with its axis of polarization oriented perpendicular to the polarizer on the illumination leg. Its purpose is to reduce specular reflections from the illuminated site. Additionally, there is a 1.3 O.D. neutral density filter (Thorlabs, Newton, NJ) in front of the lens in order to improve visualization of speckles on the sensitive camera (i.e. avoid saturation). This optical path is folded by inserting a silver mirror (Thorlabs, Newton, NJ) after the polarizer to achieve a compact layout of the system. Modeling in Zemax 13 (Zemax Inc., Kirkland, WA) showed that this configuration results in a minimum detectable speckle size that is roughly twice the camera pixel size, matching the sampling criterion for performing LSCI [24]. In the NIRAF path, the longer wavelength fluorescence is transmitted through the dichroic mirror and is further filtered by a ~800 nm longpass filter (Semrock, Rochester, NY) before being focused onto a second near-infrared-optimized camera operated at 12-bit depth. Lastly, there is a focus tunable lens (Optotune, Dietikon, Switzerland) in front of the assembly, to enable fine adjustments in focus for images acquired during surgery. The field of view of the imaging system at the working distance of 400 mm is approximately 26 × 32 mm. This platform is computer controlled, via a custom image acquisition and analysis program created in LabVIEW 2017 (National Instruments, Austin, TX).

2.2. Imaging procedure & parameters

Transportation of the imaging system over time results in slight shifts in the positions of the two cameras relative to each other. Therefore, there is a need to quantify the degree of image overlap between the cameras are prior to each imaging session. This was achieved by imaging an irregular grid and using these images to determine the rigid transformation that aligns the fields of view of the two cameras together. An efficient solution was engineered through the control software to acquire these images and subsequently perform intensity-based image registration to determine this transformation, with a single click. Fluorescence images are generated by first capturing a background image with the laser off, then a second image with the laser on, and then subtracting the former from the latter. Similar to image registration, this is executed by a single click. Due to the comparatively long exposure time required for parathyroid NIRAF, background-subtracted autofluorescence images are only generated as stills. Speckle contrast images are generated using the spatial method with a 5-by-5 pixel window, and displayed in real time at ~24 fps. The first 30 speckle contrast images after generating a fluorescence image are saved for post-processing. Sensor integration times were 100 ms and 5 ms for NIRAF and LSCI, respectively.

2.3. Fluorescence sensitivity & resolution measurement

Parathyroid autofluorescence is weak in intensity, and so to evaluate the capability of the imaging system to detect it, dilute solutions of ICG were made and imaged. Diluting with water, concentrations of 1, 0.7, 0.5, 0.3, 0.2, 0.1, 0.07, 0.05, 0.03, 0.02, and 0.01 μg/mL were made. Clinically, 3.5 mL of a 2.5 mg/mL ICG solution is administered to assess parathyroid vascularity [20]. Assuming 5 L of blood in the human body, and given that ICG remains entirely within the blood until it is broken down by hepatic cells, the above-listed concentrations represent signal levels about 2 to 200 times smaller than would be expected for intraoperative imaging of ICG fluorescence. The solutions were pipetted into two 24-well plates, with one empty well in all directions separating each sample in order to prevent signal contamination. The last well was filled with water. Fluorescence images were acquired and the average intensity across each well was calculated from these images.

To measure spatial resolution, light from an 810 nm LED was collimated onto a negative 1951 United States Air Force resolution test target (Thorlabs, Newton, NJ). The negative test target transmits light only through the engraved patterns, allowing this setup to result in the lines mimicking fluorescent sources. An image was acquired of the test target from the opposite end of the light source. The image of the resolution test target is shown in Figure 2.A. The largest set of unresolvable horizontal and vertical lines indicates the resolving power of the system. This was element 1 of group 4 (Figure 2.C), indicating a spatial resolution of 62.5 μm.

FIGURE 2.

Resolution of the NIRAF imaging system measured by imaging a negative 1951 USAF resolution test target (A) placed in front of a collimated 810 nm beam. Comparing element 6 of group 3 (B) with element 1 of group 4 (C) shows that the latter is the largest set of unresolvable horizontal and vertical lines. This corresponds to a spatial resolution of 62.5 μm.

2.4. Development of parathyroid segmentation algorithm

Previous work using LSCI to assess parathyroid vascularity required manual segmentation of the images in order to determine a parathyroid gland’s speckle contrast [23]. This is time-consuming and a potential barrier to clinical translation, and therefore would benefit greatly from a method to automate the segmentation. Combining LSCI with NIRAF imaging allows the parathyroid gland to be automatically segmented in the fluorescence image (where it is generally the most fluorescent organ). This information can then be used to automate determination of the parathyroid speckle contrast. Following is a description of the algorithm developed to automatically segment the parathyroid gland after acquiring autofluorescence and speckle contrast images. The first step is to crop the autofluorescence image (Figure 3.A) to remove the first 200 pixels on each edge (original image size is 1024 × 1280 pixels). Given that the parathyroid gland is expected to be in the center of the image, this step reduces the size of the data and speeds up computation for later steps. This image is then thresholded into three levels using a multiple thresholding scheme based on Otsu’s method [25] (Figure 3.B). The reason for choosing three levels is that, while the parathyroid is generally the strongest auto-fluorescing tissue in the neck at this wavelength, other tissues such as the thyroid also emit fluorescence; choosing three levels allows separation of parathyroid, thyroid (and other less fluorescent tissues), and the non-fluorescent background. After thresholding, the middle intensity level (thyroid) is set equal to the low intensity background (Figure 3.C). By this point, there should ideally be a clear distinction between the tissue of interest and everything else, however noise due to specular reflections and imperfect thresholding may still contaminate the image. To reduce these effects, the dominant cluster of high intensities is located through an iterative process of finding the centroid of the high intensity pixels and removing outliers to a two-dimensional Gaussian centered on the centroid until the mean and covariance of the pixel coordinates converge (Figure 3.D). This filtered image is then converted to an edge map (Figure 3.E) using Canny edge detection [26], and an active contour model is fit to this edge map to obtain the contour demarcating the parathyroid gland (Figure 3.F). The force used to drive the active contour towards the parathyroid boundary is based on gradient vector flow [27]. After obtaining the contour, the transformation that aligns the two cameras (determined from imaging the irregular grid) is applied to demarcate the parathyroid in the first of the acquired series of speckle contrast images. Nine subsequent speckle contrast images are registered to the first using a discrete Fourier transform registration that only accounts for translation [28] – since these images are acquired at 24 fps, there is not much motion from frame to frame and a transformation that only relies on translation is sufficient to register them. These ten images are then averaged to improve spatial resolution and the average speckle contrast of the parathyroid (area within the transformed contour) is obtained. The entire process described above takes about 5 seconds to run on the system computer.

FIGURE 3.

Parathyroid localization steps: background-subtracted fluorescence image (A) is thresholded into 3 levels (B). The second intensity level is set equal to the background (C), and then the dominant cluster of high intensities is located (D). An edge map is created from this cluster of points (E) and used to fit an active contour model to demarcate the parathyroid (F). The image shown is of a healthy parathyroid gland in a parathyroidectomy case.

2.5. Evaluation of segmentation algorithm

To test the device and algorithm, two thyroid and five parathyroid ex vivo tissue samples, all from different patients, were imaged in the laboratory. The specimens were obtained from the Cooperative Health Tissue Network at the Vanderbilt University Medical Center. The parathyroid specimens were obtained from cases of parathyroid adenoma, while the thyroid specimens were from non-malignant thyroid tissue. One thyroid specimen was obtained from cauterized tissue and was therefore more optically absorbing than the other. Each parathyroid specimen was placed next to each thyroid specimen for imaging, resulting in a total of ten pairings. The tissue samples were placed in a flat container with just enough phosphate-buffered saline to keep them hydrated without submerging them. Autofluorescence images were acquired of each pairing and the automated parathyroid segmentation algorithm was then run on these images. Blinded from the segmentation results, three different users manually segmented the autofluorescence images to demarcate the parathyroid. The performance of the algorithm was evaluated by calculating the Hausdorff distances between the contour generated by the algorithm and the contours manually segmented by the three users. The Hausdorff distance is the longest of all the distances from a point on one contour to its closest point on the other contour [29]. Calculations were performed for both the forward and backward direction and the maximum value was taken. Hausdorff distances between the manually segmented contours were also calculated. Two sample t-tests were performed on the two sets of data to determine if the Hausdorff distances from the contours produced by the algorithm to the manually segmented contours were significantly different from the inter-user Hausdorff distances.

Finally, a simulation was performed to evaluate the impacts of (1) parathyroid to thyroid fluorescence ratio, and (2) parathyroid signal-to-background ratio, on the segmentation algorithm. An image of a parathyroid specimen next to a thyroid specimen was simulated by manually drawing the specimens onto a blank figure in MATLAB (The MathWorks Inc., Natick, MA) and assigning intensities to the three different portions of the image: parathyroid, thyroid and background. The intensity of the thyroid was fixed, while the intensities of the background and parathyroid were varied to achieve signal to background ratios of 5, 3.5, 3 and 2, and parathyroid to thyroid fluorescence ratios of 1, 1.1, 1.2, 1.5 and 2. Normally-distributed random noise was generated with standard deviation equal to 10% of the thyroid intensity and added to each image prior to running the segmentation algorithm. Success or failure was determined by whether or not the resulting contour bounded just the simulated parathyroid.

2.6. Intraoperative imaging

One patient undergoing thyroid lobectomy and two patients undergoing parathyroidectomy at Vanderbilt University Medical Center were recruited under a larger study approved by the Institutional Review Board, and written informed consent was obtained prior to their participation. In the thyroid lobectomy case, speckle contrast images of both the superior and inferior parathyroid glands were acquired after resection of the thyroid lobe. Following this, 1 mL of a 2.5 mg/mL ICG solution was administered to the patient and fluorescence images were acquired for comparison with the speckle contrast data. In the parathyroidectomy cases, imaging followed one of two protocols depending on the health of the gland being imaged. For a diseased gland, one set of images (a single autofluorescence image and a series of speckle contrast images) was first acquired after the surgeon located and exposed the gland. Then, the surgeon ligated the blood supply to the diseased parathyroid in preparation for excision and a second set of images was acquired. For a healthy gland, only one set of autofluorescence and speckle contrast images was acquired and this occurred at any point during the surgery at the discretion of the surgeon. No more than 5 minutes was added to each surgical procedure due to imaging. Imaging was performed with the room lights on, but with the overhead OR lamps and the surgeon’s headlamp pointed away from the surgical field to avoid saturating the cameras.

3. RESULTS

3.1. Fluorescence sensitivity

Figure 4.A shows the fluorescence images obtained of ICG solutions ranging in concentration from 0.01 μg/mL to 1 μg/mL. Water is also shown (0 μg/mL). The average intensity across the channels appears to follow an exponential curve (Figure 4.B). Figure 4.C is a log-log plot of just the ICG data (excluding water), showing a good fit to a straight line with R² value of 0.98. As mentioned previously, the concentrations measured range from 2 to 200 times smaller than the expected ICG concentration intraoperatively.

FIGURE 4.

Fluorescence images of ICG solutions ranging in concentration from 1 μg/mL down to 0 μg/mL (water) in wells (A). The average intensity across each well was calculated and appeared to follow an exponential pattern (B). A logarithmic plot of the data fits a straight line with R² value of 0.98 (C).

3.2. Evaluation of segmentation algorithm

The segmentation algorithm was evaluated on autofluorescence images acquired from ten different pairings of ex vivo thyroid and parathyroid tissue samples, shown in Figure 5. The unique parathyroid specimens are sorted along the columns in order of increasing fluorescence intensity, while the two thyroid specimens are separated by the rows. The results of the automated parathyroid segmentation algorithm are indicated by the dotted cyan contours. Visually assessing the results, the algorithm correctly identified the parathyroid gland in each of the pairings. Hausdorff distances, the longest of all the distances from a point on one contour to its closest point on the other contour, were calculated in pixel numbers and converted to millimeters using the known dimensions of the field of view. The maximum inter-user Hausdorff distance in all ten pairings was 1.5 mm, while the maximum Hausdorff distance between the results of the algorithm and any user was 1.7 mm. For reference, a normal parathyroid gland is up to 8 mm in length [8]. In general, the brighter the parathyroid sample, the more similar the segmentations were to each other and therefore the smaller the Hausdorff distances. Grouping the Hausdorff distances into inter-user, and between the algorithm and each user, two sample t-tests were performed (results shown in Figure 6). Only one significant difference was found (p < 0.05) and this was for the dimmest parathyroid specimen: while the users generally conformed to each other with an average inter-user Hausdorff distance of 0.6 mm, the algorithm generated a significantly different contour resulting in the maximum Hausdorff distance of 1.7 mm. A closer look at the top left panel of Figure 3 shows that a bright spot close to the top edge of the parathyroid may have had a large influence in pulling the active contour further away from the edge, resulting in the disparity.

FIGURE 5.

Autofluorescence images of ex vivo parathyroid (para) and thyroid (thy) tissue samples. Five parathyroid samples are sorted along the columns in order of increasing fluorescence intensity, each paired with two thyroid samples separated by the rows. Results of automated parathyroid segmentation are shown by the dotted cyan contour.

FIGURE 6.

Hausdorff distances between parathyroid contours manually segmented by three different users (Inter-user), and between contours produced by the algorithm and their corresponding manually segmented contours (Users to algorithm). Data is sorted according to parathyroid fluorescence intensity normalized to the dimmest parathyroid and data from each unique parathyroid is separated from the others with dashed vertical lines. Hausdorff distances decrease the brighter the parathyroid. There is no significant difference in segmentation results except in the case of the dimmest parathyroid sample.

Simulation results (Figure 7) showed that at a parathyroid signal to background ratio of 5, the algorithm correctly segments out a parathyroid with a fluorescence intensity as low as 1.1 times that of the thyroid. At lower signal to background ratios, success of the algorithm requires stronger parathyroid fluorescence as would be expected. In the case of a parathyroid signal to background ratio of 2, success was only achieved when the background and thyroid intensities were the same.

FIGURE 7.

Results of varying parathyroid signal to background ratio (SNR) and parathyroid to thyroid fluorescence ratio on a simulated image. At higher SNR, successful segmentation is achieved at parathyroid to thyroid fluorescence ratios as low as 1.1. Stronger parathyroid fluorescence is needed at lower SNR.

3.3. Intraoperative images

Two parathyroid glands were identified by the surgeon during the thyroid lobectomy case, after removal of the diseased lobe. The surgeon considered the superior gland to be very well vascularized, and the inferior gland to be damaged but not completely devascularized as it did not appear visually dark. The average speckle contrast (calculated after manual segmentation) of the superior gland was 0.072, and that of the inferior gland was much higher at 0.128, indicating worse perfusion. The ICG fluorescence images supported this data, with the superior gland exhibiting very strong ICG fluorescence while the inferior gland had very little. These results are shown in Figure 8.

FIGURE 8.

White light, speckle contrast and ICG fluorescence images of a well-vascularized left superior parathyroid gland and a partially devascularized left inferior parathyroid gland, taken during a thyroid lobectomy case. The low speckle contrast of the superior gland is supported by strong ICG fluorescence in the gland, and the high speckle contrast of the inferior gland is supported by its minimal ICG fluorescence. Parathyroid glands are indicated by white ellipses.

A healthy parathyroid gland observed during parathyroidectomy is shown in Figure 9. Strong autofluorescence of the gland facilitated successful segmentation (visually assessed) by the algorithm. The average speckle contrast within the transformed contour in the speckle contrast image was 0.175. For comparison, the value was 0.172 within the contour produced by manual segmentation. In Figure 10, a diseased parathyroid gland is shown before and immediately after ligation of its blood supply. Visually, there is no appreciable change in either white light or autofluorescence images from before to after blood supply ligation. The only noticeable change is in speckle contrast which increased as expected, indicating a decrease in blood flow. Figure 10 highlights the heterogeneity in autofluorescence of some diseased parathyroid glands. This heterogeneity caused the segmentation algorithm to demarcate the more fluorescent portion of the gland, rather than its entirety. The resulting average speckle contrast values obtained from these contours were 0.072 before blood supply ligation and 0.188 after ligation. The uniformity of speckle contrast across the gland however, resulted in these values not being too different from the results of manual segmentation: 0.066 before and 0.186 after blood supply ligation. The larger difference (0.072 minus 0.066) was only 5% of the change in speckle contrast from pre- to post-ligation.

FIGURE 9.

White light (A), autofluorescence (B), and speckle contrast (C) images of a healthy parathyroid gland observed in a parathyroidectomy case. The dotted cyan contour in the autofluorescence image is the result of the segmentation algorithm, and the dotted white contour in the speckle contrast image is the transformed contour within which average speckle contrast is calculated.

FIGURE 10.

White light, autofluorescence and speckle contrast images of a diseased parathyroid gland before and immediately after blood supply ligation. As expected, the only appreciable change is the increase in speckle contrast, indicating a decrease in blood flow. Heterogeneity of autofluorescence resulted in a sub-total segmentation of the parathyroid gland, however speckle contrast values within the resulting contours were still similar to manual segmentation.

4. DISCUSSION

It is crucial during thyroid and parathyroid surgical procedures for the surgeon to be able to not only identify but also assess the state of vascularity of parathyroid glands. This work presents a device that performs both functions without the need for exogenous contrast agents. Near-infrared autofluorescence imaging has been shown in numerous studies to be capable of identifying parathyroid glands [9–17], and this was demonstrated in this work as well. Other methods of parathyroid identification include preoperative sestamibi scans [30], frozen section histology [31], intraoperative methylene blue administration [32], and optical coherence tomography [33]. In comparison to these, NIRAF imaging has advantages in that it is label-free, can be done in real time, and does not require expensive or complex instrumentation. On its own, the use of LSCI to assess parathyroid gland vascularity would require input from the surgeon or another member of the surgical team to identify the region of interest. By combining NIRAF imaging with LSCI, the fluorescence of the parathyroid can be used to automatically determine its location in the speckle contrast images and enable more direct relay of information to the surgeon, reducing a barrier to clinical translation. The ParaSPAI achieves this.

Given that parathyroid NIRAF is a relatively weak signal, it was important to evaluate the capability of the ParaSPAI to detect this signal. To this end, ICG solutions that were 2 to 200 times less concentrated than the levels expected after a standard administration were imaged. The fit to the data in Figure 4.C shows that the device is sensitive to this range of fluorescence intensities. For comparison, ongoing work suggests that parathyroid NIRAF is 10 to 100 smaller than the ICG fluorescence from the same gland after the dye has been injected (unpublished data). This indicates that the device should be sensitive to parathyroid NIRAF. Tiny fluorescent sources were mimicked by collimating 810 nm light through a 1951 USAF resolution test target. The resolution of the imaging system (62.5 μm) was well below the normal size of parathyroid glands, indicating that size should not hinder the detection of the parathyroid NIRAF.

In this work, we demonstrated that LSCI performs similar to ICG angiography in the same thyroid lobectomy case (Figure 8). Both methods agreed with the surgeon’s assessment that the left superior parathyroid gland was very well vascularized, and the left inferior gland was at least partially devascularized. In comparison with ICG angiography, using LSCI to assess parathyroid gland vascularity has a few advantages: the technique can be performed simultaneously with autofluorescence imaging, is more readily quantifiable, and does not require exogenous contrast. Another potential advantage is in avoiding certain false positive situations that may arise. With a half-life of about 3.4 minutes [20], it will take about 17 minutes for a bolus of ICG to washout to 3% of its initial concentration. A scenario can be envisioned in which a parathyroid gland becomes devascularized after initially demonstrating strong ICG fluorescence, due to additional surgery being performed. This gland would likely still exhibit strong ICG fluorescence, while LSCI would be able to detect the devascularization as it is sensitive to the flow of blood. On the other hand, ICG angiography has an advantage over LSCI in that it can delineate the arterial blood supply to the parathyroid gland, allowing the surgeon to avoid injury to it. While this is technically also possible with LSCI, the blood vessel would have to be very superficial.

In the ParaSPAI, a three-level thresholding scheme and active contour model were used to automatically segment the parathyroid. While in some cases, thresholding alone (followed by locating the dominant cluster of high intensities) would be sufficient to determine the pixels in the image containing parathyroid tissue, the additional step of fitting an active contour model to the parathyroid boundary helps overcome limitations in obtaining accurate speckle contrast averages across the gland due to noise and autofluorescence heterogeneity. The contours in Figure 10 did not encompass the entire parathyroid gland due to heterogeneity, however a larger area was still segmented than would have been with just thresholding. An alternative segmentation approach that might overcome this problem would be to use machine learning, which has been employed to segment features in real time during laparoscopic surgical procedures [34]. While this could be a future approach, there is currently insufficient data from the ParaSPAI to train such a model.

The mechanism for localizing the parathyroid in fluorescence images was found to not differ substantially from manual segmentation. Hausdorff distances are a measure of the similarity of two segmentations – the lower the value, the more similar the two contours in question are to each other. The results of segmenting NIRAF images of ex vivo tissue specimens showed that the stronger the parathyroid fluorescence, the more similar the segmentation contours produced manually and by the algorithm. The largest difference was between the algorithm and one user, producing a Hausdorff distance of 1.7 mm for a parathyroid tissue sample 8 mm in length. This was the least fluorescent parathyroid specimen, 1.5 times as fluorescent as the thyroid tissue next to it. More importantly however, with the exception of this one case, the Hausdorff distances between the algorithm and users did not differ significantly from the distances between the users. In other words, the segmentation produced by the algorithm was equivalent to having a fourth user manually segment the images. The algorithm is however highly dependent on the ratio of parathyroid to thyroid fluorescence, as well as the level of background noise. The smallest parathyroid to thyroid fluorescence ratio among the ex vivo tissue specimens imaged was 1.5. However, in studies using a fiber optic probe to measure parathyroid gland fluorescence, it was reported that this ratio can range anywhere from 1.2 to above 20[11]. Simulations show that the segmentation algorithm can localize a parathyroid gland that is only 1.1 times as fluorescent as the thyroid, provided the background noise level is low enough. From the values tested, this occurred when the background noise was 5 times smaller than the parathyroid fluorescence intensity. When the background noise was 3.5 times smaller than the parathyroid fluorescence intensity, segmentation was successful at a parathyroid to thyroid fluorescence ratio of 1.2. The background noise level intraoperatively will differ from case to case due to factors such as the other tissues present in the field of view, and the intensity and type of room lighting in the operating room as the filters employed in this instrument do not completely eliminate their contribution.

The ParaSPAI could be useful in both parathyroidectomies and thyroidectomies. In parathyroidectomies, its main utility would be in confirming the diseased gland(s) as parathyroid and identifying and assessing vascularity of any normal parathyroid glands encountered. For this reason, the subtotal segmentation in Figure 10 is acceptable as autofluorescence heterogeneity occurs in diseased parathyroid glands while healthy parathyroid glands tend to be more homogeneous in autofluorescence [16], as can be seen in Figure 9. Such a device would be even more useful in thyroidectomies however, where healthy parathyroid glands are more at risk of being accidentally excised or devascularized. Following is a description of its envisioned use in a thyroidectomy. During thyroid resection, autofluorescence images can be acquired to confirm parathyroid candidates identified by the surgeon. After the thyroid has been removed, the device could once again be used to image previously identified parathyroid glands and any new candidates. With the push of a button, the segmentation algorithm would localize the parathyroid in the fluorescence image and its corresponding location in the speckle contrast image would be used to determine its average speckle contrast value. This information can then be conveyed to the surgeon in some form (e.g. as a binary output or a probability of parathyroid devascularization). Work is currently ongoing to establish clinically relevant criteria for assessing parathyroid vascularity after thyroidectomy using LSCI.

In summary, we present ParaSPAI, a device capable of label-free parathyroid gland identification and vascularity assessment through the combination of NIRAF imaging with LSCI. This device overcomes limitations associated with the use of ICG angiography to assess parathyroid vascularity. The parathyroid segmentation algorithm developed for use with the device proved to be equivalent to manual segmentation and allows automated determination of parathyroid speckle contrast to simplify use of the device in the operating room. Such a device has the potential to help reduce hypoparathyroidism and hypocalcemia arising from complications in thyroid surgery.

Abbreviations:

ICG

indocyanine green

LSCI

laser speckle contrast imaging

NIRAF

near-infrared autofluorescence

ParaSPAI

parathyroid speckle and autofluorescence imager

DATA AVAILABILITY STATEMENT

The data that support the findings of this study are available from the corresponding author upon reasonable request.