differentiation between nerve and adipose tissue using wide-band (350-1,830 nm) in...

Lasers in Surgery and Medicine 46:538–545 (2014)

Differentiation Between Nerve and Adipose Tissue UsingWide-Band (350–1,830nm) in vivo Diffuse ReflectanceSpectroscopy

Rutger M. Schols, MD,1,2� Mark ter Laan, MD, PhD,3 Laurents P.S. Stassen, MD, PhD,1

Nicole D. Bouvy, MD, PhD,1 Arjen Amelink, PhD,2 Fokko P. Wieringa, PhD,2 and Lejla Alic, PhD2

1Department of Surgery, Maastricht University Medical Center & NUTRIM School for Nutrition, Toxicology andMetabolism, Maastricht University, Maastricht, The Netherlands2van’t Hoff Program on Medical Photonics, Netherlands Organization for Applied Scientific Research TNO, Eindhoven,The Netherlands3Neurosurgical Center Nijmegen, Radboud University Nijmegen Medical Center & Canisius Wilhelmina-Hospital,Nijmegen, The Netherlands

Background: Intraoperative nerve localization is of greatimportance in surgery. In certain procedures, wherenerves show visual resemblance to surrounding adiposetissue, this can be particularly challenging for the humaneye. An example of such a delicate procedure is thyroid andparathyroid surgery, where iatrogenic injury of therecurrent laryngeal nerve can result in transient orpermanent vocal problems (0.5–2.0% reported incidence).A camera system, enabling nerve-specific image enhance-ment, would be useful in preventing such complications.This might be realized with hyperspectral camera technol-ogy using silicon (Si) or indium gallium arsenide (InGaAs)sensor chips.Methods: As a first step towards such a camera, weevaluated the performance of diffuse reflectance spectros-copy by analysing spectra collected during 18 thyroid andparathyroid resections. We assessed the contrast informa-tion present in two different spectral ranges, for respec-tively Si and InGaAs sensors. Two hundred fifty three invivo, wide-band diffuse reflectance spectra (350–1,830nmrange, 1 nm resolution) were acquired on 52 tissue spots,including nerve (n¼22), muscle (n¼ 12), and adiposetissue (n¼ 18). We extracted 36 features from thesespectroscopic data: 18 gradients and 18 amplitude differ-ences at predefined points in the tissue spectra. Bestdistinctive feature combinations were established usingbinary logistic regression. Classification performance wasevaluated in a cross-validation (CV) approach by leave-one-out (LOO). To generalize nerve recognition applicabil-ity, we performed a train–test (TT) validation using thethyroid and parathyroid surgery data for training pur-poses and carpal tunnel release surgery data (10 nervespots and 5 adipose spots) for classification purposes.Results: For combinations of two distinctive spectralfeatures, LOO revealed an accuracy of respectively 78% forSi-sensors and 95% for InGaAs-sensors. TT revealedaccuracies of respectively 67% and 100%.Conclusions: Using diffuse reflectance spectroscopy wehave identified that InGaAs sensors are better suited for

automated discrimination between nerves and surround-ing adipose tissue than Si sensors. Lasers Surg. Med.46:538–545, 2014. � 2014 Wiley Periodicals, Inc.

Key words: diffuse reflectance spectroscopy; tissuespectral analysis; automated nerve detection; recurrentlaryngeal nerve; median nerve; adipose tissue

INTRODUCTION

The ability to visually distinguish vital anatomy, such asnerve tissue, is of great importance during all surgicalprocedures. There is a wide variety of procedures with arealistic chance of intraoperative nerve injury that mayresult in temporary or permanent dysfunction of motor orsensory nerves. Extra caution is, for example, requiredduring complicated surgical procedures like thyroidectomy[1] and total mesorectal excision [2,3], but also during lessdifficult procedures such as inguinal hernia repair [4].When spatial perception from direct sight and hapticfeedback from direct touch are lacking (e.g., duringminimally invasive surgery) nerve identification can beeven more challenging than during delicate opensurgery. Therefore, a reliable tool to enhance the contrastof nerve tissue from its surroundings is desirable forimproved intraoperative nerve detection and preservation.

Conflict of Interest Disclosures: All authors have completedand submitted the ICMJE Form for Disclosure of PotentialConflicts of Interest and have disclosed the following: [Nocompeting financial interests exist].

�Correspondence to: Rutger M. Schols, MD, Department ofSurgery, Maastricht University Medical Center, P. Debyelaan25, 6229 HX Maastricht, The Netherlands.E-mail: [email protected]

Accepted 15 May 2014Published online 4 June 2014 in Wiley Online Library(wileyonlinelibrary.com).DOI 10.1002/lsm.22264

� 2014 Wiley Periodicals, Inc.

Exploring optical spectroscopy techniques might offer aroadmap towards such a tool.Aerospace science combines hyperspectral camera tech-

nology, with pre-acquired library spectra recorded on theearth surface, to generate satellite images for discoveringplaces of interest, for example, agricultural purposes [5,6]and military and homeland security applications [7].Furthermore, hyperspectral imaging incorporates poten-tial to facilitate image-guided surgery [8]. It has, forexample, been investigated for non-invasive intra-operative assessment of renal oxygenation (i.e., tissueoxygen saturation) during partial nephrectomy [9,10], forintraoperative enhancement of biliary imaging (i.e.,anatomical imaging) during laparoscopic cholecystecto-my [11] and for intraoperative assessment of resectionmargins for residual tumor tissue (i.e., tumor detection) inbreast cancer surgery [12].Arrays of charge-coupled devices (CCD) and comple-

mentary metal oxide semiconductors (CMOS) are the mostcommonly used detectors (camera chips) in medicalhyperspectral imaging systems, which can be composedof silicon (Si) and indium gallium arsenide (InGaAs)sensors. Si sensors cover the wavelength range of 400–1,000nm, whereas InGaAs sensors are typically sensitivein the 900–1,700nm wavelength region, and depending onthe amount of Indium-doping the longer wavelengthboundary can shift up to 2,500nm [8].Intraoperative recurrent laryngeal nerve (RLN) identi-

fication before removal of the thyroid gland is of greatimportance. The RLNdiameter is on average 2mm [13]. Ina retrospective analysis of 5,104 primary and 685secondary thyroidectomies, transient vocal problems asa consequence of RLN palsy occurred in respectively 2%and 1% of cases (permanent in 0.5% and 1.5%). Further-more the rate of permanent complications was found to besignificantly higher in reoperative surgery [14]. Routinevisual RLN identification currently remains the goldstandard for preventing iatrogenic nerve injury duringthyroid surgery [15]. Intraoperative EMG-based nervemonitoring (IONM) [16], has been introduced to improvethe incidence of RLN palsy. However, so far no significantbenefit of IONM above visual inspection alone has beenfound [17].Optical spectroscopy techniques have shown potential

for differentiation of biological tissues as a basis for afeedback system to enhance nerve preservation in oraland maxillofacial surgery [18]. Ex vivo [18,19] and invivo [20] spectroscopic measurements (350–650 nm) havebeen performed on skin, fat, muscle, and nerve tissuesduring animal experiments. ROC analysis (after leave-one-out cross-validation) showed that almost all tissuetypes could be differentiated very well (accuracies of upto 100%) by diffuse reflectance spectroscopy, followedby principal component analysis, and linear discriminantanalysis. However, this study was performed on alimited number of animals. From ex vivo and in vivoexperiments in a swine model, Nachab�e et al. [21,22]showed the benefit of extending the ultraviolet andvisible wavelength range into the infrared, up to

1,600 nm, providing additional information regardingtissue concentrations of the biological chromophoreswater and lipid. Cao et al. [23] underlined the potentialof multispectral imaging in the so-called extendednear-infrared window based upon these endogenouschromophores.

The present study is a first fiber-optic wide band (350–1,830nm) in vivo human spectrometric exploration, withthe ultimate long-term goal of obtaining nerve-specificimage enhancement during surgery. We use signal-intensity independent features derived from spectroscopicdata acquired during thyroid and parathyroid surgery. Toexplore the robustness of the classification, we tested theperformance using the spectral data from nerve tissue in adifferent anatomical region, acquired during carpal tunnelrelease surgery.

METHODS

In vivo human tissue measurements were performed atthe Department of Surgery of Maastricht UniversityMedical Center (MUMC, Maastricht, The Netherlands)during thyroid and parathyroid surgery (surgeon N.D.B.)and at the Department of Neurosurgery of CanisiusWilhelmina Hospital (CWZ, Nijmegen, The Netherlands)during carpal tunnel release procedures (surgeon M.t.L.).Prior to measurements, approvals were granted by thelocal institutional review boards of Maastricht UniversityMedical Center (registration numberMETC 10-4-035) andRadboud University Nijmegen Medical Center (registra-tion number 2012/446), and preoperative written informedconsent was obtained from all patients.

Material

Diffuse reflectance spectra were acquired using customdeveloped sterile disposable optical fiber probes (TNO,Eindhoven the Netherlands & Light Guide Optics,Rheinbach, Germany), a modified Xenon light source (D-light C, Karl Storz, Tuttlingen, Germany), and a spec-trometer (Analytical Spectral Devices, Inc., Boulder, CO)equipped with two sensor technologies: a silicon (Si) basedsensor and an indium gallium arsenide (InGaAs) basedsensor (cross-over point at 1,000nm). The fiber probetransports the light from the light source to the tissue andfrom the tissue to the spectrometer. The spectrometeracquires the spectral data in the range of 350–1,830nmwith 1nm spectral resolution. The systemwas installed ona compact trolley to facilitate in vivo tissue measurementsduring routine surgery. The setupwas tested and approvedaccording to the essential requirements of IEC 60601-1 toassure patient safety. The equipment previously has beendescribed in more detail [24].

Figure 1 shows the fiber probe tip composed of eightsmall optical fibers: one central receiving fiber (400mmdiameter, NA 0.22� 0.02) and seven illuminating fibers(300mm diameter, NA 0.22�0.02). The rigid stainlesssteel probe tip has a 2mm diameter and a length of10mm. The remaining length of the sterile fiber probe isflexible.

AUTOMATED DISCRIMINATION BETWEEN NERVE AND ADIPOSE TISSUE 539

In Vivo Data Acquisition

During respectively thyroid and parathyroid surgeryand carpal tunnel release surgery, in vivo wide-banddiffuse reflection spectra were collected. For each tissuetype, we recorded five spectra per site (taking 30 secondsper site) covering at least one site per tissue type (see alsoTable 1). The sterile fiber probe was handled by thesurgeon and gently brought into direct contact with one ofthe designated tissues (see Fig. 1). If blood was visiblypresent on the tissue surface, it was dapped away using asterile gauze. Between the measurements on differentlocations, the probe tip was swiped with a clean sterilegauze wetted with saline. Acquired data was labeledaccording to the tissue type description of the attendingsurgeon. To correct for dark current [8], the spectrometerwas calibrated prior to in vivo data acquisition. After thecompletion of in vivo spectroscopy, a reference spectrumwas acquired, for calibration purposes, by direct contactmeasurement on a white reference phantom (Optical-grade spectralon reference; Labsphere, Inc., NorthSutton, NH). The integration times of the Silicon andInGaAs sensor were individually optimized during theSpectralon calibration. No correction for ambient lightwas performed.

Fig. 1. Fiber probe data acquisition during thyroid surgery.Intraoperative spectral fiber probe measurement performedduring a left hemithyroidectomy procedure. The sterile opticalfiber probe is gently brought into contact with the recurrentlaryngeal nerve (average diameter 2mm [13]). Right lower corner:close-up of the fiber probe tip. A ring light of seven fibers(illuminated white) injects light into the tissue, one central fiber(here for illustration purposes illuminated yellow) collects thediffusely reflected light.

TABLE 1. Spectroscopy in Thyroid and Parathyroid Surgery: Study Subject Characteristics

Subject# Gender Age Indication for surgery Surgical procedure

Sites per tissue

Nerve Adipose Muscle

1 M 67 Bethesda 3�node left thyroid Left HT 2 0 0

2 F 55 PH Right PAR 0 2 0

3 M 54 MEN-2a syndrome, left

medullar thyroid carcinoma

TTþ radical LND neck, right axillary

LND

2 1 0

4 F 14 PH Left PAR 0 2 0

5 F 67 Status after right

hemithyroidectomy, in which

a follicular carcinoma was

found

Resection of remaining thyroid (left HT) 2 1 0

6 F 53 Multinodular goiter TT 2 1 0

7 F 56 Bethesda 2�node right thyroid Right HT 2 1 0

8 F 67 Multinodular goiter Left HT 1 0 0

9 F 55 Multinodular goiter Left HTþ subtotal right thyroidectomy 1 0 0


11 M 80 PH Left PAR 0 1 1

12 F 54 PH Right PAR 1 1 1

13 M 74 PH Right PAR 0 1 1


15 F 46 Substernalmultinodular goiter Sternotomyþ left HT 2 1 3

16 F 53 Bethesda 3�node right thyroid Right HT 2 2 2

17 F 23 Graves’ disease TT 2 1 1


Sites per tissue type (total n¼52) 22 18 12

Spectra per tissue type (total n¼253) 108 90 55

M, male; F, female; PH, primary hyperparathyroidism; PAR, parathyroid adenoma resection; HT, hemithyroidectomy; TT, totalthyroidectomy; LND, lymph node dissection.��The Bethesda system was used for reporting thyroid cytopathology [29].

540 SCHOLS ET AL.

Data Processing

For inter-patient comparability, all raw in vivo spectrawere corrected using the dark current and referencereflectance spectrum. Since both reflectance intensity andspectral shape are related to the composition of the tissue,no further normalization steps were performed. To identifypossible distinctive features for tissue-specific enhance-ment, 36 features (i.e., 18 gradients and 18 amplitudedifferences at predefined points in the tissue spectra)were extracted based on known wavelengths related tocharacteristic absorption features for blood, water andfat [25–27]. Gradients are “slopes” between two predefinedpoints in the tissue spectra, that is: (DR2–DR1)/(l2–l1).Amplitude differences are “intensity differences” betweentwo predefined spectral points, that is, (DR2–DR1). DR¼diffuse reflectance; l¼wavelength. Figure 2 illustrates thecharacteristic wavelengths and features in a mean spec-trum for adipose tissue data. All data processing was

performed by in-house developed software (usingMATLABenvironment Version 7.7.0, MathWorks, Inc., Natick, MA).

Further analysis of the in vivo spectra was divided intotwo wavelength response ranges: Si-sensor (�1,000nm)and InGaAs-sensor (�900nm).

The number of extracted features (n¼36) is too large toperform a statistically meaningful classification, as theextracted features could be redundant in the informationthey retain. Therefore, using combinations of all 36features to build a classifier would result in a dimension-ality problem and over-fitting. We identified the mostdistinctive features, for classification of nerve in an adiposesurrounding, by using binary logistic regression (SPSS,Inc., Chicago, IL) for both wavelength regions separately.This is a statistical technique that allows the prediction ofcategorical dependent variables (here the tissue type:nerve or adipose tissue) using a set of independentvariables (here the spectral features) [28].

Using the approach for tissue classification based onhyperspectral data reported earlier byAkbari et al. [29], weused support vector machine (SVM) to classify nervewithin adipose surrounding. We used a polynomial kernelfunction [30] for both wavelength regions. The SVMclassifier attempts to find an optimum line in the two-dimensional feature space, consisting of support vectors, toseparate the training data with a minimum risk [31]. Toestimate classification performance and to prevent overlyoptimistic results [18–20], we implemented a cross-validation (CV) approach. With a goal to obtain theclassification accuracy as a performance measure, thedata set is divided into a training set (to train the classifier)and a test set (to validate the classifier). First the CVapproach uses leave-one-out (LOO) validation of nerve andadipose hyperspectral data acquired during thyroid andparathyroid surgery. This approach utilizes the same dataset for both training and testing purposes and is veryuseful in cases of a relatively small data sample.Additionally the CV approach uses train–test (TT) valida-tion. In this approach we divided the data into a train set(consisting of nerve and adipose hyperspectral dataacquired during thyroid and parathyroid surgery) and atest set (consisting of nerve and adipose hyperspectral dataacquired during carpal tunnel release surgery). TTvalidation provides additional information as it estimatesthe performance of this system in the clinical setting byexpanding the validation to other anatomical sites.

Sensitivity, specificity, positive predictive value, nega-tive predictive value, and accuracy were calculated toquantify the classification performance for both wave-length regions (i.e., Si-sensor and InGaAs-sensor detectionrange), and for both cross-validation methods (LOO andTT). In-house developed classifiers (using MATLABenvironment Version 7.7.0, MathWorks, Inc.) were usedto estimate the classification performance.

RESULTS

In 18 patients (4 male, 14 female) undergoing parathy-roid or thyroid surgery, 253 spectra from 52 in vivo tissuesites were recorded (nerve¼ 22, muscle¼ 12, and adipose

Fig. 2. Example of spectrum (adipose tissue neck) with charac-teristic wavelengths and investigated features. The arrowsindicate qualitative landmarks for: oxygenated/reduced hemoglo-bin (characteristic “W-shape” between 500 and 600nm), water(absorption peaks at 965 and 1,440nm) and human fat (absorptionpeaks at 1,210 and 1,720nm) [25–27]. The spectral detectionranges of the human eye, as well as for the Si-sensor and InGaAs-sensor from the applied spectrometer are also indicated. Bn, bloodrelated; Wn, water related; Fn, fat related; Ft, feature.


tissue¼18). Table 1 summarizes patient characteristicsand the number of measured sites per tissue type. Meandiffuse reflectance spectra and corresponding standarddeviations for RLN, surrounding adipose tissue andsternocleidomastoid muscle are shown in Figure 3A.Figure 3B shows mean spectra of tissue measurementsperformed during five carpal tunnel release procedures.During these measurements 20 in vivo tissue sites werestudied (nerve¼ 10, muscle¼ 5, and adipose tissue¼5).

Classification of Spectral Data in Si-Sensor Range

From the 36 extracted features, defined in Figure 2, 11were located within the silicon detection range. Given thestudy sample of 22 RLN spots, inclusion of maximum twofeatures is allowed. Binary logistic regression identifiedgradient Ft5 (B1–W1) and amplitude difference Ft7 (B5–B1)as most promising combination for differentiation of RLNfrom surrounding adipose tissue. Figure 4 shows a scatterplot for these Si-sensor based features, extracted for nerve,

and adipose tissue. Data for both thyroid and parathyroidsurgery, and carpal tunnel release surgery are included.The quantitative results of classification performance

are listed in Table 2 for both cross-validation approaches.LOO cross-validation is solely based on Si-sensor rangedata from thyroid and parathyroid surgery for train andtest purposes. TT cross-validation is based on data fromthyroid and parathyroid surgery for train purposes and onadditional carpal tunnel release surgery data for testpurposes.

Classification of Spectral Data in InGaAs-SensorRange

From the 36 extracted features, defined in Figure 2, 25were locatedwithin the spectral detection range of InGaAs.After binary logistic regression, gradient Ft12 (W1–F4) andamplitude difference Ft26 (F2–F1) were selected as themost promising combination for differentiation of RLNfrom surrounding adipose tissue. Figure 5 shows a scatterplot for these InGaAs-sensor based features extracted fornerve and adipose tissue. Data for thyroid and parathyroidsurgery and carpal tunnel release surgery data areincluded.The classification performance for respectively LOO

cross-validation and TT cross-validation are listed inTable 3. LOO cross-validation is solely based on InGaAs-sensor range data from thyroid and parathyroid surgeryfor train and test purposes. TT cross-validation is based ondata from thyroid and parathyroid surgery for train

Fig. 3. A: Mean spectra per tissue type acquired during thyroidand parathyroid surgery. Average tissue spectra for recurrentlaryngeal nerve (green), adipose tissue (blue), and sternocleido-mastoid muscle (red). Dashed lines in the corresponding colorsindicate the respective standard deviations. RLN diameter is onaverage 2mm [13] at the measurement site. B: Mean spectra pertissue type acquired during carpal tunnel release surgery.Average tissue spectra for median nerve (green), subcutaneousadipose tissue (blue), and transverse thenar/hypothenar musclefibers overlying the transverse carpal ligament (red). Dashed linesin the corresponding colors indicate the respective standarddeviations. Median nerve diameter is on average 6mm at themeasurement site (i.e., at distal fore-arm [40]).

Fig. 4. Scatter plot of two selected features within Si-range.Scatter plot showing two computer-selected features (gradient Ft5and amplitude difference Ft7). Data measured during thyroid andparathyroid surgery and carpal tunnel release surgery areincluded.

542 SCHOLS ET AL.

purposes and on additional carpal tunnel release surgerydata for test purposes.

DISCUSSION AND CONCLUSION

This explorative study reveals the in vivo wide-band(350–1,830nm) diffuse reflectance spectra of humanrecurrent laryngeal nerve, surrounding adipose tissueand sternocleidomastoidmuscle (anatomical region: neck).Also in vivo spectra are presented for human mediannerve, subcutaneous adipose tissue, and transversethenar/hypothenar muscle fibers overlying the transversecarpal ligament (anatomical region: wrist). The presentedspectra covered both silicon (Si) and indium galliumarsenide (InGaAs) sensor ranges (350–1,830nm with

1nm resolution), thereby exceeding beyond the 1,600nmboundary reported by preceding work [21–23,32]. Spec-troscopy on human skin samples in the wavelength rangeof 1,000–2,200nm has been reported [33]. However, thisreport did not include any of the tissues covered by ourstudy.

Even though themeasured reflectance spectra, shown inFigure 3, illustrate a great similarity between differenttissue types, the nerve classification within an adiposesurrounding was fairly accurate within the detectionranges of both Si and InGaAs. Based on the classificationaccuracies for these detection ranges (Tables 2 and 3) wecan conclude that InGaAs sensors are better suited forautomated discrimination between nerves and surround-ing adipose tissue than Si sensors.

With respect to the statistically different spectralfeatures of nerve versus adipose tissue in the InGaAsrange, our results indicate that especially water and lipidabsorption differences around 1,210nm play an importantrole. This is also strongly supported by previously reporteddata on composition of nerve and adipose tissue [23,34,35].In essence, nerve tissue contains 20% lipid and 80%protein [34], while adipose tissue contains 61–87% lipid,8% protein and 11–31% water [35].

The reflectance spectra, which were the basis to extractgradient and amplitude difference features, originate fromintrinsic tissue properties (endogenous contrasts) that donot require preoperative contrast administration. Conse-quently there are no problems with potential toxicity orallergy to a contrast agent.

Regarding exogenous contrast-based optical techniques,in vivo optical imaging of peripheral nerves usingsystemically administered myelin-selective fluorescentdyes [36] or nerve-highlighting fluorescent peptides [37]has been reported. Development of a new NIR fluorescentdye for use in the design of nerve-targeted optical imagingprobes has also been described [38].

The LOO cross-validation method inherently producesrelatively optimistic classification results (therefore weadditionally applied the TT cross-validation method).External validation remains essential before classificationmodels can be implemented in clinical practice [39]. Suchvalidation would need to be performed on newly acquireddata. Additional data acquired in a multi-center studywould also be needed.

TABLE 2. Classification Performance of Selected Si-Sensor Features

TP TN Sensitivity Specificity PPV NPV Accuracy

LOO CV 20/22 11/18 91 (69–98) 61 (36–82) 74 (53–88) 85 (54–97) 78

TT CV 10/10 0/5 100 (66–100) 0 (0–54) 67 (39–87) — 67

TP, true positive; TN, true negative!numbers indicate identified tissue spots. A positive test is defined as the tissue observed beingRLN; a negative test is defined as the tissue observed being adipose tissue.Sensitivity; specificity; PPV, positive predictive value;NPV, negative predictive value; accuracy!numbers are percentages; numbers inparentheses indicate 95% confidence interval.LOO CV, leave-one-out cross-validation; TT CV, train–test cross-validation.

Fig. 5. Scatter plot of two selected features within InGaAs-range.Scatter plot showing two computer-selected features (gradientFt12 and amplitude difference Ft26). Data measured duringthyroid and parathyroid surgery and carpal tunnel releasesurgery are included.


The choice to use a limited set of pre-known spectralregions of interest for water, fat, and hemoglobin wasmadebased upon the limited number of spectra included in ourdata set. Differentiation based on completely automaticallyextracted features from a larger data set might achievebetter results, and could be explored in future work.

The identified reflectance spectra are specific to theprobe geometry used in this study. To extend this studyeven further, translation of these reflectance spectra to invivo biological parameters is needed. Such an approachwould yield information about the optical tissue properties,and provide a better understanding of the nature ofdiscrimination performance, that is, whether nerves can beoptically distinguished from surrounding (adipose) tissuesbased on differences in light scattering behaviour (relatedto structural differences) or due to differences in absorp-tion behaviour (related to chromophore concentrationssuch as blood, water and fat). Such an approach would alsobe tissue-specific and robust to inter-patient and multi-centre variability.

The “gold standard” used in this study was the surgeons’visual judgment. This judgment is not solely based on color(spectral) information, but also relies on the recognition ofspatial anatomical position of a specific tissue. Therefore,this study needs extension from spot-wise probe-measure-ments to imaging of the whole surgical field.

In contrast to our previous ex vivo experiments usingwide band diffuse reflectance spectroscopy [24], thecurrent measurements were performed during surgery.In vivo circumstances as vascular filling and oxygenationwere not disrupted. Using diffuse reflectance spectroscopywe have identified that InGaAs sensors are bettersuited for automated discrimination between nerves andsurrounding adipose tissue than Si sensors. This under-lines the importance of extending the wavelength regionbeyond the detection boundary of silicon cameras forautomated classification of nerve tissue in adiposesurroundings. Hyperspectral camera technology coveringboth Si and InGaAs sensor response ranges couldoffer even greater possibilities for automated tissuedifferentiation.

ACKNOWLEDGEMENTS

We would like to thank all involved medical and ORstaff and the Department of Medical Technology fromMaastricht University Medical Center and Canisius

Wilhelmina Hospital for helping us to perform the clinicalmeasurements.

REFERENCES

1. Christou N, Mathonnet M. Complications after total thyroid-ectomy. J Visc Surg 2013;150:249–256.

2. Lange MM, Maas CP, Marijnen CA, Wiggers T, Rutten HJ,Kranenbarg EK, van de Velde CJ. Urinary dysfunction afterrectal cancer treatment is mainly caused by surgery. Br JSurg 2008;95:1020–1028.

3. Lange MM, van de Velde CJ. Urinary and sexual dysfunctionafter rectal cancer treatment. Nat Rev Urol 2011;8:51–57.

4. Lange JF, Wijsmuller AR, van Geldere D, Simons MP, SwartR, Oomen J, Kleinrensink GJ, Jeekel J. Feasibility study ofthree-nerve-recognizing Lichtenstein procedure for inguinalhernia. Br J Surg 2009;96:1210–1214.

5. Huang W, Lamb DW, Niu Z, Zhang Y, Liu L, Wang L.Identification of yellow rust in wheat using in-situ spectralreflectance measurements and airborne hyperspectral imag-ing. Precision Agric 2007;8:187–197.

6. Hively WD, McCarty GW, Reeves JB, Lang MW, OesterlingRA, Delwiche SR. Use of airborne hyperspectral imagery tomap soil properties in tilled agricultural fields. Appl EnvironSoil Sci 2011;2011:13 pages. Article ID 358193, doi: 10.1155/2011/358193

7. Yuen PWT, Richardson M. An introduction to hyperspectralimaging and its application for security, surveillance andtarget acquisition. Imaging Sci J 2010;58:241–253.

8. Lu G, Fei B. Medical hyperspectral imaging: A review. JBiomed Opt 2014;19:10901.

9. Holzer MS, Best SL, Jackson N, Thapa A, Raj GV, CadedduJA, ZuzakKJ.Assessment of renal oxygenation during partialnephrectomy using hyperspectral imaging. J Urol 2011;186:400–4404.

10. Olweny EO, Faddegon S, Best SL, Jackson N, Wehner EF,Tan YK, Zuzak KJ, Cadeddu JA. Renal oxygenation duringrobot-assisted laparoscopic partial nephrectomy: Characteri-zation using laparoscopic digital light processing hyper-spectral imaging. J Endourol 2013;27:265–269.

11. Zuzak KJ, Naik SC, Alexandrakis G, Hawkins D, BehbehaniK, Livingston E. Intraoperative bile duct visualization usingnear-infrared hyperspectral video imaging. Am J Surg 2008;195:491–497.

12. Panasyuk SV, Yang S, Faller DV, Ngo D, Lew RA, FreemanJE, Rogers AE. Medical hyperspectral imaging to facilitateresidual tumor identification during surgery. Cancer BiolTher 2007;6:439–446.

13. Serpell JW, Woodruff S, Bailey M, Grodski S, Yeung M.Recurrent laryngeal nerve diameter increases duringthyroidectomy. Ann Surg Oncol 2011;18:1742–1747.

14. Lefevre JH, Tresallet C, Leenhardt L, Jublanc C, Chigot JP,Menegaux F. Reoperative surgery for thyroid disease.Langenbecks Arch Surg 2007;392:685–691.

15. Chiang FY, Wang LF, Huang YF, Lee KW, Kuo WR.Recurrent laryngeal nerve palsy after thyroidectomy withroutine identification of the recurrent laryngeal nerve.Surgery 2005;137:342–347.

TABLE 3. Classification Performance of Selected InGaAs-Sensor Features

TP TN Sensitivity Specificity PPV NPV Accuracy

LOO CV 22/22 16/18 100 (82–100) 89 (64–98) 92 (72–99) 100 (76–100) 95

TT CV 10/10 5/5 100 (66–100) 100 (46–100) 100 (66–100) 100 (46–100) 100

TP, true positive; TN, true negative!numbers indicate identified tissue spots. A positive test is defined as the tissue observed beingnerve; a negative test is defined as the tissue observed being adipose tissue.Sensitivity; specificity; PPV, positive predictive value;NPV, negative predictive value; accuracy!numbers are percentages; numbers inparentheses indicate 95% confidence interval.LOO CV, leave-one-out cross-validation; TT CV, train–test cross-validation.

544 SCHOLS ET AL.

16. Dralle H, Sekulla C, Lorenz K, Brauckhoff M, MachensA. Intraoperative monitoring of the recurrent laryngealnerve in thyroid surgery. World J Surg 2008;32:1358–1366.

17. Pisanu A, Porceddu G, Podda M, Cois A, Uccheddu A.Systematic review with meta-analysis of studies comparingintraoperative neuromonitoring of recurrent laryngeal nervesversus visualization alone during thyroidectomy. J Surg Res2014;188:152–161.

18. Stelzle F, Zam A, Adler W, Tangermann-Gerk K, Douplik A,Nkenke E, Schmidt M. Optical nerve detection by diffusereflectance spectroscopy for feedback controlled oral andmaxillofacial laser surgery. J Transl Med 2011;9:20.

19. Stelzle F, Tangermann-Gerk K, Adler W, Zam A, Schmidt M,Douplik A, Nkenke E. Diffuse reflectance spectroscopy foroptical soft tissue differentiation as remote feedback controlfor tissue-specific laser surgery. Lasers Surg Med 2010;42:319–325.

20. Stelzle F, Adler W, Zam A, Tangermann-Gerk K, Knipfer C,Douplik A, Schmidt M, Nkenke E. In vivo optical tissuedifferentiation by diffuse reflectance spectroscopy: Prelimi-nary results for tissue-specific laser surgery. Surg Innov2012;19:385–393.

21. Nachabe R, Hendriks BH, Desjardins AE, van der Voort M,van der Mark MB, Sterenborg HJ. Estimation of lipid andwater concentrations in scattering media with diffuse opticalspectroscopy from 900 to 1,600nm. J Biomed Opt 2010;15:037015.

22. Nachabe R, Hendriks BH, van der Voort M, Desjardins AE,Sterenborg HJ. Estimation of biological chromophores usingdiffuse optical spectroscopy: Benefit of extending the UV–VISwavelength range to include 1000 to 1600nm. Biomed OptExpress 2010;1:1432–1442.

23. Cao Q, Zhegalova NG, Wang ST, Akers WJ, Berezin MY.Multispectral imaging in the extended near-infrared windowbased on endogenous chromophores. J Biomed Opt 2013;18:101318.

24. Schols RM,DuniasP,WieringaFP, StassenLP.Multispectralcharacterization of tissues encountered during laparoscopiccolorectal surgery. Med Eng Phys 2013;35:1044–1050.

25. Zijlstra WG, Buursma A, Meeuwsen-van der Roest WP.Absorption spectra of human fetal and adult oxyhemoglobin,de-oxyhemoglobin, carboxyhemoglobin, and methemoglobin.Clin Chem 1991;37:1633–1638.

26. Anderson RR, Farinelli W, Laubach H, Manstein D,Yaroslavsky AN, Gubeli J 3rd, Jordan K, Neil GR,Shinn M, Chandler W, Williams GP, Benson SV, DouglasDR,DyllaHF. Selective photothermolysis of lipid-rich tissues:A free electron laser study. Lasers Surg Med 2006;38:913–919.

27. RogganA, FriebelM,DoRschelK,HahnA,MuLlerG.Opticalproperties of circulating human blood in the wavelengthrange 400–2500nm. J Biomed Opt 1999;4:36–46.

28. Tabachnick BG, Fidell LS. Using multivariate statistics, 5thedition. Boston, MA: Allyn-Bacon; 2007.

29. Akbari H, Kosugi Y, Kojima K, Tanaka N. Blood vesseldetection and artery–vein differentiation using hyperspectralimaging. Conf Proc IEEE Eng Med Biol Soc 2009;2009:1461–1464.

30. Mercier G, Lennon M. Support vector machines for hyper-spectral image classification with spectral-based kernels.Geoscience and Remote Sensing Symposium, 2003 IGARSS’03 Proceedings 2003 IEEE International 2003;1:288–290.

31. Boser BE, Guyon IM, Vapnik VN. 1992; A training algorithmfor optimal margin classifiers. Proceedings of the 5th AnnualACM Workshop on Computational Learning Theory: 144–152.

32. Desjardins AE, van der Voort M, Roggeveen S, Lucassen G,Bierhoff W, Hendriks BH, Brynolf M, Holmstrom B. Needlestylet with integrated optical fibers for spectroscopic contrastduring peripheral nerve blocks. J Biomed Opt 2011;16:077004.

33. Troy TL, Thennadil SN. Optical properties of human skinin the near infrared wavelength range of 1000 to 2200nm.J Biomed Opt 2001;6:167–176.

34. Malmivuo J, Plonsey R. Bioelectromagnetism: Principles andapplications of bioelectric and biomagnetic fields, 1st edition.Oxford: Oxford University Press; 1995.

35. WoodardHQ,WhiteDR. The composition of body tissues. Br JRadiol 1986;59:1209–1218.

36. Gibbs-Strauss SL, Nasr KA, Fish KM, Khullar O, Ashitate Y,Siclovan TM, Johnson BF, Barnhardt NE, Tan Hehir CA,Frangioni JV. Nerve-highlighting fluorescent contrast agentsfor image-guided surgery. Mol Imaging 2011;10:91–101.

37. Whitney MA, Crisp JL, Nguyen LT, Friedman B, Gross LA,Steinbach P, Tsien RY, Nguyen QT. Fluorescent peptideshighlight peripheral nerves during surgery in mice. NatBiotechnol 2011;29:352–356.

38. Gustafson TP, Yan Y, Newton P, Hunter DA, Achilefu S,Akers WJ, Mackinnon SE, Johnson PJ, Berezin MY. A NIRdye for development of peripheral nerve targeted probes.MedChemComm 2012;3:685–690.

39. Bleeker SE, Moll HA, Steyerberg EW, Donders AR, Derksen-Lubsen G, Grobbee DE, Moons KG. External validation isnecessary in prediction research: A clinical example. J ClinEpidemiol 2003;56:826–832.

40. Mani B, Sarawagi R, Cherian RA. Review of the dimensions ofthe median nerve and carpal tunnel using sonography inasymptomatic adults. J Med Imaging Radiat Oncol 2011;55:126–131.


differentiation between nerve and adipose tissue using wide-band (350-1,830 nm) in...

Documents