2.1.

Study Design

The overall study comprised, first, measurements of the Raman spectra of bone and soft tissues in freshly excised swine vertebra ex vivo. Those measurements were used as input data to a supervised machine learning support vector machine (SVM) analysis that associated each tissue type with the corresponding Raman spectral fingerprint. This resulted in a trained and validated multi-class predictive model designed to discriminate between six tissue types: bone, cartilage, fat, ligament, muscle, and spinal cord (model I). The ex vivo measurements were also used to produce two-class models trained and validated to discriminate between bone and spinal cord (model II) and between bone and all soft tissue types (Model III). The predictive accuracy (sensitivity and specificity) of all tissue classifiers was tested using an independent hold-out dataset that was not used during the model training/validation phase.

To emulate real-world conditions met during orthopedic procedures, the two-class models were independently tested on two other Raman spectral fingerprint datasets. One set was acquired in situ from freshly excised vertebra, while the other was acquired in vivo under conditions closer to the real-world surgical scenario. In both cases, the spectroscopic data were acquired during bone drilling using the RS probe.

2.2.

Raman Spectroscopy System and Measurement Protocol

The system used to obtain tissue spectral fingerprints (Sentry 1000-R, Reveal Surgical Inc., Montreal, Canada) is an NIR RS instrument that has been approved for investigational clinical testing in neurosurgery by Health Canada. Mounted on a portable cart, it consists of illumination and detection modules, a hand-held probe, and a laptop computer (Fig. 1).

Fig. 1

RS system. (a) Photograph of the reveal system with major components annotated: (I) laptop with graphical user interface, (II) illumination and detection module housing, (III) hospital-grade isolation transformer, (IV) laser and spectrometer fiberoptic connection panel, and (V) fiberoptic bundle with optical connectors and hand-held probe. (b) Close up of the probe in contact with a vertebral tissue sample. (c) Example Raman spectrum.

The illumination module houses a 785 nm diode laser (class IIIB) with a maximum output of 100 mW (Innovative Photonic Solutions, Plainsboro, New Jersey). The detection module comprises an NIR spectrometer and associated optical and electronic components. The spectrometer consists of a charge-coupled device sensor (Newton model, Andor Technology, Belfast, United Kingdom) cooled to $-40°\mathrm{C}$, a $100\mu \mathrm{m}$ spectrometer slit, and a diffraction transmission holographic grating. The probe has a central light-excitation fiber surrounded by nine collection fibers ($100\mu \mathrm{m}$ core diameter) potted within a stainless-steel ferule and terminated with a tapered tip and interface window. The probe is connected to the laser and the spectrometer through a 3 m long fiberoptic cable. It is sterilizable, reusable, and has the shape of a 12 cm long stylet. Where the probe contacts tissue, there is a conical tip of outer diameter 2.1 mm. Optical filters are mounted within the probe tip to minimize signals from the fiber materials and the tissue autofluorescence. The spectrometer has high sensitivity across the range 400 to $2000{\mathrm{cm}}^{-1}$, with an average resolution of $1.8{\mathrm{cm}}^{-1}$. A converging lens at the tip of the probe ensures contact measurements interrogated a 0.5 mm diameter spot. The system is controlled by proprietary software (Reveal Surgical Inc., Montreal, Canada) that allows acquisition parameters to be set by the user, including laser power, exposure time per spectrum, and number of repeated measurements (i.e., accumulations) at each point. A laboratory version of the system has been used in several clinical studies for intraoperative tumor resection guidance in various organs including breast,³⁴ brain,^24,35,36 ovaries,³⁷ and prostate.^23,38–40

System calibration comprised acquisition on an NIST Raman standard (SRM 2214) to correct for the instrument response.⁴¹ The probe tip was cleaned with isopropyl alcohol before and after calibration, as well as throughout the study to minimize cross-contamination of signals from previously measured tissues. For each spectral measurement, a suitable illumination exposure time was selected based on the optical properties of the tissue under investigation to ensure maximum photon counts without saturating the detection electronics. Once a measurement site was selected, the tip of the probe was held in contact with the tissue at a right angle to maximize optical coupling [Fig. 1(c)]. Two individuals participated in the measurements, one preparing the animal model and tissues and the other operating the system, thereby minimizing potential motion artifacts during each measurement. All measurements were carried out under low ambient lighting to minimize background noise. At least three spectral fingerprints were collected at each measurement location (i.e., accumulations), using total integration times ranging from 0.4 to 20 s per spectrum, depending on the tissue type. For example, typical integration times per accumulation for bone and spinal cord were 1.0 and 0.8 s, respectively.

2.3.

Animal Protocol

All animal studies were carried out under institutional review and approval (AUP #6578: University Health Network, Toronto, Canada). To minimize the number of swine needed to meet the study objectives, animals that were already enrolled in other approved terminal studies were used for the ex vivo and in situ measurements. The in vivo study was performed only after these prior data demonstrated high and consistent results. The animals were free from genetic manipulation, tracer probes or any other exogenous substances except anesthesia. The 33 kg male Yorkshire swine was fasted 8 h prior to surgery and received meloxican IM ($0.4\mathrm{mg}/\mathrm{kg}$) preoperatively. General anesthesia was induced and maintained with inhaled isoflurane, 4% and 2.5%, respectively. During the entire procedure, the animal had all vital signs monitored closely and received buprenorphine IM ($0.4\mathrm{mg}/\mathrm{kg}$) and fluids i.v. After the procedure, the animal was euthanized by intravenous injection of KCl ($150\mathrm{mg}/\mathrm{kg}$).

2.4.

Ex Vivo Spectroscopic Measurements

The spinal columns of three swine were cut with a surgical bone saw to remove one thoracic vertebra, along with intact soft tissues that interface with the bone. Figure 2 is an example, showing the cortical and trabecular bone (henceforth lumped together into the “bone” category) and soft tissues (cartilage, fat, ligament, muscle, and spinal cord). Prior to collection of the Raman spectra, the sample was lightly rinsed with saline to help identify the structures visually and minimize spectral crosstalk from blood and cerebrospinal fluids. On average, eight different measurement sites were selected per tissue type on each vertebra, contingent on availability of suitable exposed tissue surfaces. The tissue class was identified based on visual inspection, with knowledge of spinal anatomy.

Fig. 2

Example of freshly excised vertebra in (a) axial and (b) lateral views, with the major components and Raman spectral fingerprints labeled.

2.5.

In Situ and In Vivo Spectroscopic Tissue Measurements

In situ spectra were acquired for each tissue category (bone, cartilage, fat, ligament, muscle, and spinal cord) in two freshly sacrificed animals. Spinal surfaces were reached using a conventionally open long-segment incision technique for pedicle access prior to drilling in the vertebra. The bone surface was exposed so that tissues could be visually identified. In total, 14 holes were drilled from four different vertebrae in each animal [Figs. 3(a)–3(c)]: eight lumbar, five thoracic, and two cervical. The last was the most difficult to access due to a thick overlaying muscle mass. Overall, this resulted in 50 spectral measurements in one animal and 38 spectra in the other.

Fig. 3

(a)–(c) In situ and (d) in vivo drilling and RS procedures. (a) Photographs at each stage of drilling demonstrating mm-level depth control and showing the bone surface and the drilling cavity at the cortical/trabecular interface (2 mm), trabecular bone (3 mm), trabecular/cortical interface (6 mm), epidural space (8 mm), and spinal cord breach (9 mm). (b) In situ drilling under fluoroscopy imaging guidance with a labeled radiograph shown in (c). (d) 3D CT image guidance in the axial view showing labeled vertebra with mis-directed drill direction toward spinal canal (red arrow), indicated by insertion of contrast-infused gauze. The green arrow indicates the correct drill trajectory through the pedicle.

Following in situ testing, in vivo measurements were performed in one anesthetized animal to confirm the system and machine-learning model performance under more clinically realistic conditions [Fig. 3(d)]. Four vertebrae in total were examined from thoracic, lumbar, and cervical sections of the spine. Since a critical requirement during spinal fusion surgery is to ensure proper placement of pedicle screws, detection of misdirected screw trajectories resulting in spinal canal breaches was a primary focus during trajectory planning. The in vivo measurements resulted in 44 spectral fingerprints acquired along four different trajectories, one trajectory per vertebra: three vertebrae had 12 spectral measurements, and one had eight measurements.

All in situ and in vivo measurements were performed by intra-pedicular drilling in 1 mm steps using a cordless drill equipped with a shaft-collar drill-stop mounted on a 6 mm titanium drill bit with a 135 deg drill point angle. Cooling and lubrication were provided by saline rinsing of the drilling cavity at each step prior to measurement of the spectral fingerprint. The drill stop provided mm-level depth control, permitting precision measurement of spectra from tissue layers and layer interfaces. In situ guidance during drilling was carried out as shown in Fig. 3(c) under x-ray fluoroscopy (OEC 9800, GE Healthcare). The in vivo drilling was guided using 3D CT (Cios Spin C-arm, Siemens Healthcare Ltd.), as shown in Fig. 3(d), where a (deliberate) spinal canal breach was indicated using contrast medium-soaked gauze pressed into the drilling cavity.

2.6.

Data Pre-Processing and Raman Spectral Fingerprint Interpretation

The following data pre-processing steps were applied to each spectroscopic measurement:⁴¹ (1) subtraction of a dark count background measurement acquired with the laser turned off prior to each repeat acquisition (e.g., to remove residual contamination from ambient light sources), (2) $x$-axis (wavenumber) normalization and instrument response correction from spectral measurements acquired in calibration materials (acetaminophen powder and NIST 785 nm Raman standard, respectively), (3) curve smoothing using a Savitzky-Golay filter of order 3 with a window size of 11 (unit-less optimization parameter), (4) averaging of successive measurements acquired at the same location, (5) baseline subtraction using a custom algorithm, BubbleFill.⁴¹ Finally, the resulting Raman spectral fingerprints were normalized based on standard-normal-variate (SNV). This normalization technique implies that the “intensity” associated with each spectral bin or band must be interpreted as a variation relative to the average of all detected inelastic scattering contributions across the spectral domain.

Recent RS studies have contributed reliable and well-characterized spectra associated with specific tissues. Raman peaks are narrow (typically $<50{\mathrm{cm}}^{-1}$) and in many cases can be associated with a specific chemical bond or functional group. For example, Movasaghi et al.²² have compiled the most frequent reported Raman bands in tissues. For each tissue type considered here (bone, cartilage, fat, ligament, muscle, and spinal cord), the main visually distinguishable Raman bands were identified on the ex vivo Raman spectral fingerprints, mostly based on the Movasaghi et al. paper. The resulting band assignments are listed in Table 1 and were also cross-checked against other publications.

Table 1

Principle vibrational modes, band assignments, and corresponding relative concentration in each tissue type indicated by the number of asterisks from lowest (*) to highest (****).22 The tissue label “connective” refers to the categories “ligament” and “cartilage” combined.

2.7.

Machine Learning Tissue Classification Models

Three different machine learning models were produced from the ex vivo measurements. Model I consisted of a six-class predictive model (bone versus cartilage versus fat versus ligament versus muscle versus spinal cord), while models II and III were two-class models. Model II differentiated bone from spinal cord tissue and model III discriminated bone from all types of soft tissues. All models were tested on a hold-out data subset: 60% for model training and validation and 40% for testing. Further, models II and III were independently tested using the datasets acquired under conditions closer to real-world, i.e., the in situ and in vivo datasets.

Each processed spectrum led to a Raman spectral fingerprint comprising more than 900 spectral bins. Prior to machine-learning model development, the spectra were reduced to $N<20$ spectral features using a linear SVM approach with $L1$-regularization (Lasso regression) optimization.⁴² The resulting dimensionally reduced features set—in the form of $N$ pre-selected spectral intensities associated with specific wavenumber values—was used to train and validate the machine-learning models using a linear SVM approach with $L2$-regularization (Ridge regression). Both phases—feature selection and machine learning modeling—were associated with one SVM hyperparameter that is conventionally labeled $C$. The values of both hyperparameters ($C_1$ for feature selection, $C_2$ for machine learning model production) were optimized through a grid search running over a large range of combinations. The $C$-parameter associated with the feature selection phase, $C_1$, was varied between 0.005 and 0.05. This ensured that the number of retained features was always $<20$ to minimize the risk of over-fitting the data during the model development phase, during which $C_2$ was varied from 0.1 to 5.

Each combination of hyperparameters $(C_1,C_2)$ led to a different machine-learning model. A training/validation process was utilized to find the parameters leading to optimal predictive performances, i.e., to select those parameters that led to the smallest number of false positive and false negative predictions. Importantly, this needed to be done ensuring the final models generalized well to new data, namely, to ensure an optimal balance was reached between under-fitting and over-fitting the spectral data. This was achieved during a training/validation phase using a fivefold cross-validation technique applied on 60% of the ex vivo dataset, retaining the remaining 40% to evaluate model performances on an independent holdout data subset.

For models II and III, the performance during the training/validation phase (using 60% of the data) was assessed through a receiver-operating-characteristic (ROC) analysis. An ROC curve ($x$-axis, specificity; $y$-axis, 1 - sensitivity) was generated and the selected final model had the largest sensitivity and specificity values. Here, this corresponded to the point on the ROC curve that had the shortest distance to the upper-left corner. The $2\times 2$ confusion matrix associated with the ROC analysis was also produced to showcase the number of false positive/negative and true positive/negative instances. The optimal machine-learning model was then applied directly onto the holdout dataset (40% of the ex vivo dataset) and the resulting predictive performances were reported using another confusion matrix.

All data acquired during the in situ and in vivo experiments were then used as another way to test the models using independent data to assess their generalizability. All data from those experiments were applied to models II and III, leading to tissue classes prediction, along with a quantitative measure of the probability of association (an output from the SVM analysis) with each class ranging from 0 to 1.³⁷ For example, in the case of model II, a probability of association close to 1 indicates that the model predicted, with high confidence, that the Raman spectral fingerprint was associated with spinal cord tissue, while a value closer to 0 indicated that the model confidently associated the measurement to bone.

ROC analyses do not lend themselves to performance analyses in the case of multi-class (i.e., more than two classes) models. Hence, the performance for Model I was assessed using a confusion matrix $M_{ij}$, where $i$ and $j$ each run from 1 to 6. The diagonal elements of the matrix report the number of correct predictions for each class, while the off-diagonal elements tabulate the number of incorrect model predictions. Specifically, off-diagonal elements $(i,j)$ with $i\ne j$ correspond to the number of predictions associating a Raman spectral fingerprint to tissue category $i$ that should have been associated with tissue category $j$, and vice-versa. As in the performance assessment of Models II and III, two confusion matrices were produced, one for the training/validation phase and one from the testing phase. For the latter, the six-class tissue discrimination model was applied directly to the hold-out set comprising 40% of the ex vivo dataset. The analysis associated with applying model I to the in situ and in vivo datasets is not presented here for conciseness.

3.1.

Band Assignment and Raman Spectral Fingerprint Interpretation

Figure 4(a) shows the complete ex vivo dataset as a spectrogram, individual SNV-normalized Raman spectra, vertically stacked and grouped by tissue type. The intensity at each wavenumber is represented by a false-color scale, ranging from dark blue (minimum) to light green (maximum). Figure 4(b) shows the average spectrum for each tissue category, together with the variance across all measurements, computed for each spectral bin. Table 1 summarizes the visually detectable Raman bands for all tissue types that are associated with known molecular vibrational modes. A tentative association is made to families of biomolecules, and the relative intensity for each peak center is shown. The relative intensity (across all tissue types) for each peak is indicated using a scale from 1 to 4 (represented by asterisks: *, **, *** or ****) based on the results of univariate statistical analysis (student $t$-test) applied to the average peak intensities. Since each of the tissues of interest has distinct molecular composition, the distinguishing Raman signatures can be clearly seen.

Fig. 4

Multi-class ex vivo tissue classification. (a) Spectrogram showing the entire ex vivo dataset, demonstrating qualitatively the consistency of specific tissue spectra across all measurements. (b) Average spectra per tissue type, with the corresponding variance computed for each spectral bin across all measurements. Vertical gray bands represent the features selected for machine-learning classification. The confusion matrices are shown in (c) and (d) for the training and testing sets, respectively.

Looking at the muscle column in Table 1, most bands were directly linked to proteins and amino acids and were in good agreement since proteins are the most important component of striated skeletal muscle.⁴³ The peaks at 1131 and 1448 to $1451{\mathrm{cm}}^{-1}$ encompass fatty acids and lipids; the presence of which is expected since intramuscular fat accumulates both within (intramyocellular) and surrounding (extramyocellular) muscle fibers. For the spinal cord the largest peak is at 1439 to $1441{\mathrm{cm}}^{-1}$ corresponding to DNA/RNA, protein (amide I), and lipids. The aggregation of these different molecular species represents the distribution of proteins and genetic information in nerve fiber bundles within the spinal cord. Lipids also often accompany proteins, since the spinal cord comprises both white and gray matter. The myelin sheath surrounding the nerve fibers that compose the spinal cord also has high lipid content (70% to 80%).⁴⁴ Interestingly, in the fat tissues, the highest peak is at $1305{\mathrm{cm}}^{-1}$, assignment to collagen. Adipocytes (fat cells) produce mainly collagen that aids cell adhesion, differentiation, and wound healing.⁴⁵ Most of the other peaks relate to lipids, which confirms the fat tissues composition.

Connective tissue in the table groups cartilage and ligament together, based on the main collagen constituent with the highest peak at $817{\mathrm{cm}}^{-1}$, as well as peaks at 857, 1067 to 1069, 1303, and 1448 to $1451{\mathrm{cm}}^{-1}$. The material strength and biological properties of articular cartilage depend heavily on its unique and extensively cross-linked extracellular collagen network and characteristic fibrillar organization that varies with tissue depth and cellular proximity. Ligaments are generally composed of ground substance, collagen (mainly types I and III) with minimal elastin fibers, i.e., the building blocks are collagen fibers. Another high peak at $1641{\mathrm{cm}}^{-1}$ is assigned to water,⁴⁶ which is an important component of articular cartilage, contributing up to 80% of wet weight. Part of this water is linked to the intrafibrillar space within the collagen.⁴⁷ Cartilage also has a high peak at 959 to $965{\mathrm{cm}}^{-1}$ for calcium hydroxyapatite corresponding to mineral calcium, deposits of which can be found in articular cartilage. Overall, connective tissue includes all the peaks selected, since the properties of connective tissue reside in the amount, type, and arrangement of abundant extracellular matrix. By contrast, the biological properties of tissues such as spinal cord, fat, or muscle depend mostly on their cellular elements.⁴⁸

Finally, bone has its highest peak at 959 to $965{\mathrm{cm}}^{-1}$, assigned to calcium hydroxyapatite. Bone comprises both a mineral (inorganic) and an organic phase. Calcium hydroxyapatite is the main component of the former, which makes up $\sim 60\%$ of the tissue.⁴⁹

3.2.

Ex Vivo Multi-Class Tissue Classification Model

Visual inspection of the individual Raman spectra [Fig. 4(a)] and their average for each tissue type [Fig. 4(b)] shows that the spectral fingerprints acquired on freshly excised specimens (i.e., the ex vivo spectral dataset) have clearly distinguishable features enabling tissue discrimination. As the mineral peak at $961{\mathrm{cm}}^{-1}$ is unique to trabecular and cortical bone, it can be easily distinguished from soft tissues. Although fat and spinal cord show similarities, they can be separated according to the phenylalanine peak at $1004{\mathrm{cm}}^{-1}$. Muscle, cartilage, and ligament tissues are also distinguishable, given that the peak around $1451{\mathrm{cm}}^{-1}$ is weak in cartilage, while the phenylalanine peak at $1004{\mathrm{cm}}^{-1}$ helps to further differentiate between muscle and ligament.

The vertical gray regions in Fig. 4(b) represent the spectral bands that were picked up by the feature-selection algorithm in building the six-class machine-learning model (Model I). A total of 13 spectral bands were required to achieve an accuracy $>96\%$ in both the training/validation and the testing. This is shown in the confusion matrices for the training/validation phase [Fig. 4(c)] and the testing phase [Fig. 4(d)], where only 3 of 98 and 2 of 64 spectra were misclassified, respectively.

3.3.

In Vivo and In Situ Testing of the Two-class Soft Tissue Detection Models

The confusion matrices from the training/validation and testing phases associated with model II (bone versus spinal cord) are shown in Fig. 5(a). The corresponding confusion matrices associated with Model III (bone versus soft tissues) are shown in Fig. 5(b).

Fig. 5

Two-class ex vivo tissue classification models. Confusion matrices quantifying model performance on training and testing sets: (a) model II – bone versus spinal cord and (b) model III – bone versus soft tissues.

The models performed with 100% sensitivity during both the training and testing phases. These two-class models were then applied to the in situ dataset (Fig. 6) and the in vivo dataset (Fig. 7) to evaluate how well they generalized to situations closer to real-world surgical orthopedic procedures. Both models II and III used only two features, namely the mineral peak at $961{\mathrm{cm}}^{-1}$ and the amide I peak at $1441{\mathrm{cm}}^{-1}$.

Fig. 6

Two-class soft tissue detection models applied to in situ data: model II (bone versus spinal cord), model III (bone versus soft tissues). Raman fingerprint spectrograms are shown for the entire in situ dataset, for each of the nine vertebrae, qualitatively demonstrating the transition from bone to spinal cord through the disappearance of the mineral peak at $961{\mathrm{cm}}^{-1}$ as the probe approaches soft tissue. The classification probability for each model and for each individual spectrum is shown in gray for model II and in red for model III. A value closer to 1 associates the measurement to bone, while a value closer to 0 associates it to soft tissue.

Fig. 7

Two-class soft tissue detection models applied to the in vivo data: model II (bone versus spinal cord) and model III (bone versus soft tissues). Raman fingerprint spectrograms are shown for the entire dataset for each of the four vertebrae, qualitatively demonstrating the transition from bone to spinal cord through the disappearance of the mineral peak at $961{\mathrm{cm}}^{-1}$ as the RS probe approaches soft tissue. The classification probabilities for each model and for each individual spectrum are shown in gray for model II and in red for model III.

Figure 6 shows the Raman spectrograms associated with all vertebrae for which in situ measurements were made. The $y$-axis in these plots represents the drilling direction, i.e., the axis along which the Raman measurements were made, starting from a bone region to anatomical areas associated with soft tissue, i.e., spinal cord. The transition from bone to spinal cord can be appreciated through visual inspection, namely the disappearance of the mineral peak at $961{\mathrm{cm}}^{-1}$ and the concomitant gradual appearance of the amide bands.

The spectrograms also consistently demonstrate a pattern of initial bone followed by a well-differentiated soft tissue signature, mostly resembling that of spinal cord. This is confirmed by the numerical values associated with the confidence-of-association to a tissue class when using model III (bone versus soft tissue). All measurements with a strong mineral peak and low intensity amide bands are associated with probabilities of associated to bone close to 1 (red bars in Fig. 6). Similar trends are observed when applying model II (bone versus spinal cord) to the same dataset (gray bars in Fig. 6). However, the numerical values associated with the probability of association to bone, although consistently larger than 0.5, are smaller than those obtained in model III. The higher confidence using model III may be due to the fact it was trained using a more heterogeneous dataset that better reflects the anatomical environment moving from bone to spinal cord.

Figure 7 shows that similar conclusions are reached when applying Models II and III to the in vivo measurements. However, these data have higher inter-measurement variances compared with the ex vivo and in situ measurements, likely due to interference from bleeding that both introduced additional materials and may have increased light attenuation even with the use of a saline flush during drilling. Despite this, the distinctive features of bone and spinal cord tissues still allow unambiguous identification of the bone–spinal cord interface.