Research ArticleCancer Diagnostics

Intraoperative Tissue Identification Using Rapid Evaporative Ionization Mass Spectrometry

See allHide authors and affiliations

Science Translational Medicine  17 Jul 2013:
Vol. 5, Issue 194, pp. 194ra93
DOI: 10.1126/scitranslmed.3005623

Abstract

Rapid evaporative ionization mass spectrometry (REIMS) is an emerging technique that allows near–real-time characterization of human tissue in vivo by analysis of the aerosol (“smoke”) released during electrosurgical dissection. The coupling of REIMS technology with electrosurgery for tissue diagnostics is known as the intelligent knife (iKnife). This study aimed to validate the technique by applying it to the analysis of fresh human tissue samples ex vivo and to demonstrate the translation to real-time use in vivo in a surgical environment. A variety of tissue samples from 302 patients were analyzed in the laboratory, resulting in 1624 cancerous and 1309 noncancerous database entries. The technology was then transferred to the operating theater, where the device was coupled to existing electrosurgical equipment to collect data during a total of 81 resections. Mass spectrometric data were analyzed using multivariate statistical methods, including principal components analysis (PCA) and linear discriminant analysis (LDA), and a spectral identification algorithm using a similar approach was implemented. The REIMS approach differentiated accurately between distinct histological and histopathological tissue types, with malignant tissues yielding chemical characteristics specific to their histopathological subtypes. Tissue identification via intraoperative REIMS matched the postoperative histological diagnosis in 100% (all 81) of the cases studied. The mass spectra reflected lipidomic profiles that varied between distinct histological tumor types and also between primary and metastatic tumors. Thus, in addition to real-time diagnostic information, the spectra provided additional information on divergent tumor biochemistry that may have mechanistic importance in cancer.

Introduction

More than 300,000 new cancer cases are diagnosed in the United Kingdom each year (1), and 1.8 million diagnostic, curative, or palliative surgical procedures are performed as part of their treatment. Surgical excision remains the gold standard of care for most solid tissue tumors. From an oncological perspective, a curative surgical intervention is determined by the complete removal of tumor tissue with an associated border of microscopically healthy tissue often referred to as the tumor “margin with clearances.” However, the functional or cosmetic success of surgery is also dependent on minimizing the excess removal of noncancerous or normal tissue. Evidence suggests that current oncosurgical techniques are frequently inadequate; 20% of breast cancer patients treated with breast-conserving surgery (that is, lumpectomy) require further surgery to clear positive margins (2). This is important because the surgical resection margin remains one of the most important prognostic factors in a large number of cancers (3, 4). Furthermore, re-excision is not always possible, such as in the case of soft tissue tumors or colon cancer, thus necessitating further adjuvant therapies.

Accurate local tumor invasion can only be assessed by means of careful multiple histological analyses in the postoperative phase. This is problematic during excision biopsy or curative resection of dysplasia or carcinoma in situ, where there is often no visible lesion to guide the surgical resection. Remarkably, there are almost no technologies in routine clinical practice to assist the surgeon in improving the accuracy of cancer tissue clearance, the most widely accepted methods being the use of intraoperative imaging modalities such as computed tomography or magnetic resonance imaging in neurosurgical (5) and complex ear, nose, and throat (6) procedures or fluorescence-guided surgery (with 5-aminolevulinic acid) in neurosurgical resections (7).

In cases of uncertainty, the removed tissue is sent to the pathology laboratory for intraoperative histological examination while the patient remains under general anesthesia as part of a process called “frozen section.” This takes time (20 to 30 min), and it is costly; only a limited number of sampling points are possible, and histology may be open to subjective interpretation, especially in cases of suboptimally prepared or non-oriented specimens/slides. Electrosurgical devices are commonly used for hemostasis, for improved accuracy of dissection, and for therapeutic purposes (Fig. 1). A major by-product of their use is the production of “smoke” from the evaporating tissue as it is being resected. Historically, surgical smoke has been considered as toxic and an irritant; therefore, it is often dispelled from the operative field. We hypothesized that this is a rich source of biological information and therefore used mass spectrometry to measure its metabolomic composition. Desorption ionization/mass spectrometry (DI-MS) methods have been optimized for biological samples within the last two decades (5, 6). However, application of DI-MS in a surgical environment has been hindered by the accompanying sample preparation procedure (79).

Fig. 1 Schema of REIMS instrumentation and data collection.

(A and B) Surgical ion source and ion transfer setups for REIMS experiments using monopolar electrosurgery (A) or bipolar electrosurgery (B). (C) Schematic of aerosol aspiration using commercially available bipolar forceps.

The recently developed rapid evaporative ionization mass spectrometry (REIMS) technique circumvents these challenges by using standard electrosurgical methods as means of converting tissue components into gas-phase ionic species amenable to mass spectrometric analysis (10). Tissue specificity of the REIMS method has been shown to be similar to that of other DI-MS methods (11). Because the technique does not require the modification of electrosurgical tools, testing in the surgical environment involves only the modification of the mass spectrometer and coupling to the surgical instruments, as previously described (12, 13). The resulting mass spectrometric profiles are highly specific to the type of tissue analyzed, allowing for tissue identification as well as characterization on a level comparable with histopathological analysis (11). This analytical coupling creates new chemical information sets that describe the tissue and its associated pathology. We have coined the term “intelligent knife” (iKnife) to describe this linked electrosurgical MS process that includes subsequent multivariate analysis for patient diagnosis.

Although we previously demonstrated the iKnife successfully in animal models (11), this study represents the first application of the technology in the human patient population. The objectives of this study were to test the REIMS-iKnife method in a surgical environment and to assess its accuracy in the detection of the tumor margin status in patients undergoing resection of brain, liver, lung, breast, or colorectal tumors. Before use in the surgical suite, we collected iKnife data ex vivo from benign and malignant gastric, colonic, hepatic, breast, lung, and brain tissue samples and, from those data, constructed a histologically authentic spectral reference library. Database and multivariate statistical tissue identification algorithms were then tested during surgery. The results presented here indicate that the envisioned application is feasible and that the REIMS-based iKnife technique provides a valid alternative to frozen section histology to enhance real-time intraoperative decision-making.

Results

Interpretation of raw spectra and statistical validation

Samples taken from surgically resected tumor specimens were initially analyzed ex vivo by REIMS for the construction of a histologically specific mass spectral database containing 1624 cancerous, 1231 healthy, and 78 benign and inflammatory bowel disease (IBD) spectra. One hundred ninety-nine ionic species in the mass spectra of various tissue types were identified using exact mass measurements with 30,000 resolution and tandem mass spectrometry in the mass/charge (m/z) range of 600 to 900. Identified compounds included phosphatidylethanolamines, phosphatidylcholines, phosphatidic acids, phosphatidylserines, phosphatidylinositols, sphingomyelins, cardiolipins, plasmalogens, and sulfatides (Fig. 2A). Further exploration of these spectra demonstrated that most of the lipid species were detected across multiple tissue types (Fig. 2, B and C); however, their distribution patterns were markedly distinct for healthy versus cancerous tissues. In line with other work, this demonstrated that it is the profile of these species rather than any specific biomarker(s) that accounts for the histological specificity (14). Compositional differences were also readily observed among the spectra of healthy tissue types, and, perhaps most importantly, spectral differences were also observed between different types of malignant tumors. Histological specificity, in principle, allows the identification of tissues by REIMS technology, given that a statistically significant number of histologically authentic spectra are available in the form of a database, as it is demonstrated in the present study.

Fig. 2 Typical REIMS spectral profiles and the structures of corresponding phospholipid classes.

(A) The basic structure of important phospholipids. R1 and R2 correspond to different acyl moieties. The exact mass and glycerophospholipid type of each species are specified below the structure, with the number of carbon atoms and double bonds in parentheses, and an indication of either loss of hydrogen or ammonia indicating how the negative ion is generated. (B and C) Human healthy liver spectrum (B) and human liver metastasis spectrum from a breast cancer patient (C) recorded on Orbitrap Discovery Fourier Transform MS in m/z 600 to 900. Mass/charge and relative intensity to m/z 885.55 are shown. The species in both tissues mostly overlap, as demonstrated in the insets in mass range 696 to 701; however, the relative ratios of characteristic species were different.

The subject-to-subject reproducibility of REIMS data was sufficiently good for the unambiguous identification of major tissue types: liver, lung, and colon (Fig. 3A). Leave-one-patient-out cross-validation calculated for the three distinct healthy tissue types shown in Fig. 3 resulted in 100% correct tissue identification. Multivariate statistical analysis determined the contribution of chemical species to the tissue-specific spectral differences. The principal components in the principal components analysis (PCA) were linear combinations of the normalized intensities of ionic species. After PCA, R2 = 34.6% (the percentage of the total variance), R2bcv = 32.2 for the first principal component; R2 = 10.7%, R2bcv = 8.9 for the second component; and R2 = 5.4%, R2bcv = 4.1 for the third component. [BCV stands for (3 × 3) bi–cross-validation according to (15).] The loading plot (Fig. 3B) shows the contribution of each ionic species in the REIMS spectra (from all tissue specimens) to the first and second principal components.

Fig. 3 Statistical analysis of healthy tissue data.

(A) Spectra were obtained from healthy liver parenchyma (13 patients), healthy alveolar lung tissue (28 patients), and healthy colon mucosa (47 patients). Each data point represents one sample (n). (B) The first two loading plots for healthy liver tissue (blue) and alveolar lung tissue (green) demonstrate the contribution of each lipid species to the first two principal components (PC). (C) Median/quartile/non-outlier minimum/non-outlier maximum box plots of four selected peaks from (B). P values, determined with Kruskal-Wallis ANOVA-type test, were adjusted using the Bonferroni method.

Assuming that the intergroup variance was larger than the intragroup variance—that is, the diversity of tissue types was responsible for the extension of data cloud in the multidimensional data space—the peaks contributing the most to the first two principal components served as tissue-specific indicators. The intensity distribution of three species (PA, PE-NH3, and PI) provided a relatively higher (loading coefficients > 0.04) contribution to the first two principal components compared to most of the other measured species (Fig. 3B). This was confirmed using univariate analysis of individual candidate metabolites (Fig. 3C): Compared with the other tissues, m/z 673.48 was a characteristic peak of alveolar lung tissue representing PA(34:1)-H; m/z 727.53 was a characteristic peak of colon mucosa representing PE(36:1)-NH3; whereas m/z 697.48 and m/z 885.55 were peaks found in all three healthy tissues, representing PE(34:2)-NH3 and PI(38:4)-H. The Kruskal-Wallis [analysis of variance (ANOVA)] test further confirmed that the differences in median intensities of these characteristic peaks were statistically significant between distinct tissue types (Bonferroni-adjusted P < 0.001) (Fig. 3C).

Ex vivo oncological analysis: Solid tumors

To establish a multivariate statistics–based spectral identification system, a large-scale database constituting histologically validated database entries (n = 2933) was constructed by the analysis of ex vivo human tissue samples (Table 1). All tissue specimens were analyzed by REIMS. Data analysis indicated that the iKnife approach could differentiate between distinct oncological tissue types and that the results were congruous with the histology. Linear discriminant analysis (LDA) was performed on REIMS data from fresh ex vivo solid tumors of the lung. After PCA, R2 = 37.9%, R2bcv = 34.2 for the first component; R2 = 11.7%, R2bcv = 6.8 for the second component; and R2 = 4.0% for the third component (Fig. 4A). Complete separation of different histologically specific data points strongly suggests that mass spectrometric identification of different lung tumors is feasible.

Table 1 Patients and associated spectral entries stored in the database.

The PCA-LDA model was created from every tissue type separately, and leave-one-patient-out cross-validation was used for every patient in the training set. Cross-validation was not feasible in the “Other” category because of insufficient number of entries corresponding to individual tissue types. The total number of patients includes the patients not uploaded to the database and used for creating test sets. The database can be obtained upon request.

View this table:
Fig. 4 Multivariate statistical analysis of malignant tumor data obtained ex vivo.

n represents the number of spectra in the training set. (A) Pseudo–three-dimensional (3D) LDA plot of different human lung tumors. (B) 3D PCA plot of healthy liver parenchyma and hepatocellular carcinoma. Spectra obtained within 1 cm of the border lines are also depicted. The arrow demonstrates a shift from healthy liver parenchyma through tumor border into tumor tissue. (C) Pseudo-3D LDA plot of two different brain metastases and two different WHO-grade gliomas in the human brain. Numbers of patients were 40, 22, and 14 for (A), (B), and (C), respectively.

This finding was further corroborated by PCA of REIMS spectra taken from ex vivo cholangiocarcinomas (malignant tumors originating from the epithelial cells of the bile duct). After PCA, R2 = 39.3%, R2bcv = 34.4 for the first component; R2 = 10.5%, R2bcv = 6.2 for the second component; and R2 = 5.4%, R2bcv = 1.6 for the third component (Fig. 4B). Data points corresponding to bulk tumor tissue separate well from the surrounding healthy liver tissue, and data points collected every 1 cm from the visible tumor indicate a gradual metabolic transition from a benign to a malignant tissue phenotype. [The British Society of Gastroenterology recommends >1 cm for liver resections (16), and therefore, this value was chosen as a reproducible sampling site.]

Ex vivo oncological analysis: Metastases

It was possible to identify the origin of metastatic tumors both in vivo and ex vivo. Figure 4C depicts the LDA of data obtained during neurosurgical tumor resection. After PCA, R2 = 21.3%, R2bcv = 15.2 for the first component; R2 = 8.4%, R2bcv = 3.5 for the second component; and R2 = 6.8% for the third component. Data were generated by removing a 0.1- to 1-mm3 piece of tissue from the tumor using a bipolar forceps connected to the mass spectrometer and subsequently evaporating the specimen above the surgical area. This approach allowed the use of relatively high power settings (40 to 60 W) without causing functional damage to cortical areas by stray electric currents and heating. The process required less than 3 s, including data analysis (time demand for individual steps) and feedback of histological information to the surgeon. These data show that astrocytomas with different World Health Organization (WHO) grades can be clearly differentiated from brain metastases (Fig. 4C). Moreover, brain metastases from disparate primary tumors as confirmed by histopathology (colon and lung) also demonstrated a specific lipidomic phenotype detectable by REIMS (Fig. 4C).

The metabolic field transition identified by REIMS (Fig. 4B) was only observed for primary tumors and not metastases. We therefore analyzed tissue obtained from liver samples containing colonic adenocarcinoma metastasis. Samples were surgically excised and immediately subjected to REIMS analysis using a monopolar electrocautery handpiece. Parallel sampling lines 1 to 2 mm apart were taken using a continuous burn, which started in the healthy region and continued through the center of the cancer, stopping for 1 s on the macroscopic border of healthy and cancerous tissue (Fig. 5A). All spectra obtained were matched and classified with a hematoxylin and eosin (H&E)–stained histopathological section (Fig. 5C). The spectra obtained from benign and malignant tissues are shown in Fig. 5B. By analyzing these data using the PCA-LDA model for healthy liver parenchyma and colonic adenocarcinoma metastasis, it was possible to project their spectral identity onto the 3D PCA model (Fig. 5D) for the prediction of histological class. This technique correctly identified 73 sampling points, with two false-negative and one false-positive results.

Fig. 5 Histologically assigned data acquisition.

(A) Human liver with colon cancer metastasis. (B) Spectrum of healthy liver parenchyma obtained more than 1 cm from the tumor border and cancerous tissue obtained from the center of the tumor. (C) H&E-stained image of the tissue region marked by a red box in (A). Scale bar, 2 mm. (D) 3D PCA plot of healthy liver parenchyma and colon metastasis spectra. The 12 red dots indicate the location of the 12 marked spectra in (A). Location 5 was identified as outlier.

All spectral databases were constructed using ex vivo tissue (with the exception of healthy brain tissue). However, the device is designed for both in vivo and ex vivo use. To assess the quality of the in vivo data acquisition, multivariate models were constructed to compare in vivo and ex vivo data sets. No separation of in vivo and ex vivo spectra could be observed (fig. S2), suggesting that the in vivo modeling is robust and representative. Leave-one-patient-out cross-validation resulted in 0% false-negative and 9.1% false-positive rates in separating healthy tissue from cancerous tissue.

Intraoperative analysis of tumor specimens in vivo

A tissue identification platform constituting the spectral database and the PCA-LDA–based algorithm was tested during surgical interventions in the operating theater. Results of the real-time intraoperative interpretation of the spectral data are summarized in Table 2. A total of 864 spectra were acquired from 81 patients and identified in all cases—that is, the spectra confirmed the result of postoperative histopathology. A sensitivity of 97.7% and a specificity of 96.5% were reached for binary classifications (cancer/healthy) of all cases. There was a low rate of both false-positive (3.5%) and false-negative (2.3%) results (Table 1). However, subclass analysis suggested that the predictive capacity of the model was less strong for primary liver and malignant lung tumors, which had a decreased specificity (<95%) and a false-positive rate of 6 to 8%. All percentages were determined by leave-one-patient-out cross-validation of the relevant data set.

Table 2

Testing PCA-LDA models for each tissue type. Test spectra collected across different tumor types and healthy tissues were created from patients not uploaded to the database. The outlying test spectra were removed from the analysis. In the test set, 3 to 5 microscans were averaged to create one spectrum, whereas the spectra uploaded to the database are created from the average of at least 10 microscans. The test set was created for testing the quasi–real-time function of the algorithm.

View this table:

Although mass spectrometric tissue identification matched postoperative histopathology in most of the cases (96.2%), there were 15 cases (table S1) where the preoperative diagnosis (based on histopathological examination of biopsy specimens) and postoperative histopathological tissue identification were conflicting. The REIMS-iKnife technique, however, was able to correctly identify the corresponding tissue in 11 cases, and correctly identify suspected malignant tumors as benign in 9 of those cases as corroborated by the postoperative definitive histological diagnosis (table S1). None of these patients received chemotherapy before surgical intervention. Of the two cases that were not considered benign, one was suspected colonic adenocarcinoma identified as Crohn’s disease by both MS and postoperative histopathology. The remaining case had an inconclusive preoperative diagnosis, thought to be a benign neoplasm of the lung; MS identified this as a squamous cell carcinoma, although the postoperative histological examination confirmed the tissue as an adenocarcinoma.

Discussion

These data provide compelling evidence that the REIMS-iKnife approach can be translated into routine clinical use in a wide range of oncosurgical procedures. It can augment current tumor diagnostics, and it has the potential to influence “on-table” decision-making and ultimately to improve oncological outcomes. Near–real-time identification of positive resection margins could also improve cosmetic and functional outcomes by minimizing surgical trauma and the unnecessary removal of healthy tissue.

Real-time surgical margin control requires the direct combination of a dissection tool and a tissue identification system. These two components are conveniently linked in case of REIMS, in a manner that provides optimal conditions both for dissection (for example, fine monopolar forced coagulation mode for precision surgery) and for mass spectrometric tissue identification (for example, to produce a sufficient amount of charged particles). From data-processing perspective, margin control is generally a two-tissue problem (healthy versus malignant), similar to the case depicted in Fig. 4B. The tissue involved in the analysis in this case is theoretically a single, healthy tissue type, and the instrument is expected to alert the surgeon when the identification results abandon the healthy tissue cloud in the multidimensional data space and move toward the tumor tissue data points. However, in case of the identification of unknown tissue types (for example, potential metastases), it is not imperative to add the tissue margin analysis function to the dissection tool—almost any electrosurgical hand tool designed for minimally invasive sampling would be suitable for that purpose. In this case, the data analysis model must be prepared to identify any healthy or diseased tissue type that could potentially be encountered by the operator during an intervention. Owing to the nature of the supervised multivariate statistical analysis methods implemented here, the identification efficiency is usually compromised by the increasing number of classes. This problem was tackled in the REIMS-iKnife system by narrowing the classification to a single organ system, which could be manually selected by the operator.

Spatial resolution of the presented method is determined by the sensitivity of the analyzer and the phospholipid content of tissues. Although needle electrode sampling allows micrometer-level spatial resolution, the evaporated material gives sufficient mass spectrometric signal-to-noise ratio at evaporation rates in the range of 20 to 250 μg/s. This range translates to fine surgical dissection at a linear rate of ~1 mm/s. Regarding absolute amounts of tissue, these values suggest that ~0.1 mm3 of tissue is sufficient to provide effective mass spectrometric information for the accurate identification of a bulk tissue profile. Although greater resolution has been demonstrated in alternative mass spectrometric imaging techniques, such as MALDI (matrix-assisted laser desorption ionization), DESI (desorption electrospray ionization), or SIMS (secondary ion mass spectrometry) (17, 18), our numbers correlate well with the maximum level of precision achievable during surgical intervention. One of the chief advantages of the proposed methodology lies in its unprecedented speed of tissue analysis via spectroscopic and chemometric processing of a surgical by-product (smoke). The entire procedure including sampling, sample transfer, chemical analysis, data processing, and audiovisual feedback takes ~0.7 to 2.5 s, depending on the configuration of the instrumentation and the surgical procedure in question; this can be compared to ~30 min for intraoperative histopathology. Near–real-time tissue analysis not only potentially shortens the whole surgical intervention but also enables the identification of thousands of chemical features of tissues during the procedure.

Intraoperative imaging methods such as ultrasound and magnetic resonance, selective labeling techniques, and spectroscopic characterization of tissues in vivo are being established as alternative methodologies to replace frozen sectioning and histology (1923). Although some of these methods have been developed to the level of routine applicability, problems with tissue selectivity have limited their application to narrow, well-defined areas (24, 25). In vivo labeling techniques are also occasionally used (20, 26), but pharmacokinetics and potential toxicity can further reduce their applicability. Infrared dyes and Raman spectroscopy have emerged as leading optical technologies, providing excellent selectivity in many different cases of solid tumors (2730). However, they provide only limited biological or chemical information, and in vivo data suggest that it lacks the sensitivity and specificity of REIMS (31).

We also found that metastatic deposits demonstrate a clear metabolic separation from the surrounding healthy tissue in contrast to primary tumors, where a surrounding transition zone could be detected. The substantial difference between the environment of primary and metastatic tumors was tentatively associated with the field cancerization theory (32)—that is, that cells surrounding primary tumors and malignant cells have been exposed to similar carcinogenic effects; thus, these histologically healthy cells are already in a precancerous state and have changes in their membrane lipid composition, which are readily detected by the REIMS-iKnife approach. Although there is no direct evidence for this assumption, a number of recent studies (33, 34) seem to support the concept of metaboplasia or “metabolic shadow” in the environment of primary tumors.

The REIMS-based mass spectrometric method of tissue identification goes beyond identifying tumor versus healthy tissue by successfully identifying different tumor grades in several cases. The bottleneck in the development of the system for the identification of a wider variety of pathology—including malignant tumors and inflammatory diseases—based on machine-learning approaches is that it requires a histologically specific mass spectral library of a good quality. The number of spectra necessary for the unambiguous identification can be properly determined only in a retrospective way (that is, when the addition of new elements does not improve classification performance further); however, it is estimated to be in the order of magnitude of a few thousand subjects for an organ system. In this sense, the iKnife mass spectrometric lipidomics approach offers a potential alternative to histopathology in all cases where rapid diagnosis is preferred, with the added capacity to provide new chemical information on a variety of benign and malignant tissue types.

The outlined technology is envisioned to be translatable to the operating theater and to endoscopy and pathology environments. Barriers to rapid dissemination and large-scale clinical testing thus far have included the current cost of instrumentation and the effort and time needed for the development of an MS-based medical device and the spectral database. The development of mass spectrometer platforms complying with regulatory expectations is currently in progress. We have developed a database of histologically validated mass spectra that is compatible with any instrument. However, collection and validation of the data are time-consuming, and in the initial stages, it will not be possible to identify every tissue type that may be encountered, for example, some very rare tumor types. The initial cost of instrumentation is expected to decrease as the technology becomes more widespread. If fully implemented into the surgical environment, this technology has potential to reduce the rate of local tumor recurrence and the cost of histopathology services and to improve the overall survival of patients.

Materials and Methods

Study design

This was a prospective multicenter observational study performed at three hospitals (1st Department of Surgery, Semmelweis University Hospital, Budapest, Hungary; Institute of Neurosurgery, Debrecen University Hospital, Debrecen, Hungary; and Institute of Surgery, Debrecen University Hospital, Debrecen, Hungary) between 1 May 2010 and 29 February 2012. Ethical approval was obtained from the Hungarian National Scientific Research Ethical Committee. The aim of the study was to assess the real-time, in vivo characterization of human tissue by mass spectrometric analysis of the aerosol released during electrosurgical dissection. In total, 393 patients were recruited who were undergoing surgery for gastric, colorectal, liver, breast, lung, and brain pathology (summarized in table S2). The range of tissue types was chosen to validate the general diagnostic capacity of the iKnife technology. The study was performed in two parts: (i) construction of tissue-specific database by systematic analysis of ex vivo specimens; (ii) intraoperative identification of vital tissues and fresh ex vivo samples. The ex vivo tissue specimens of 302 patients were initially used for the construction of the tissue-specific MS database. The sample size of the database was chosen to ensure the stability of classification via “learning curve fitting” (35), that is, the specimens were collected until no further gain in the tissue classification accuracy with the increased training sample size (database entries) was observed. The cross-validated statistical models constructed for mass spectrometry–based tissue identification were subsequently validated for intraoperative prediction of tumor status in 81 patients.

Instrumentation

The experimental setup used throughout the study is depicted in Fig. 1 and fig. S1A. Details of the mass spectrometers and electrosurgical handpieces are in the Supplementary Methods.

Construction of histologically assigned spectral database

The workflow for the acquisition of database entries is illustrated in fig. S1. All specimens were collected in theater blocks of the Institute of Surgery and the Institute of Neurosurgery of Debrecen University and the 1st Department of Surgery of Semmelweis University between 1 May 2010 and 29 February 2012. Bulk tumor samples were cut into macroscopic slices. One of the freshly cut surfaces was analyzed with monopolar REIMS using a sharp needle electrode. Data collection was performed with fresh ex vivo tumor specimens. The remaining part of the tumor specimen was formalin-fixed and embedded into paraffin, and the cut surface was sectioned for histological examination. Sections were stained by H&E (or subjected to immunohistochemical staining in some cases) and scanned with a panoramic microscope slide scanner (3DHISTECH) at a sufficient magnification (×40 to ×100) to allow for visual comparison to the mass spectrometric image

Intraoperative REIMS

For the intraoperative studies, the instrument was installed in a general surgery operating room (fig. S1A). The handpiece and aerosol aspiration line are shown in fig. S1B. Mass spectra were recorded intraoperatively by surgeons whenever electrosurgery was used for dissection or coagulation. Excised tumor tissue was also sampled with the electrosurgical device after removal from the body (<10 s). Unknown tissue features were point-sampled in vivo, in situ. Localization of the origin of mass spectra was performed by taking continuous video shot of the intervention, synchronized with the mass spectrometric data acquisition. Resected tissue samples underwent detailed histopathological examination, and the result of the histological examination of the tissue along the electrosurgical dissection line was compared to the results of mass spectrometric tissue identification (an example is shown in fig. S1, C and D).

Data acquisition and statistical analysis

The sampled tissue yielded charged molecular species, which were subjected to separation based on differing mass-to-charge ratios and subsequent quantitative detection within the mass spectrometer. Because the REIMS method generally produces molecules with single charge, the recorded signal for any detected molecular species corresponds only to its molecular mass, simplifying data interpretation. The separation and detection of charged species require 150 to 500 ms per iteration, and therefore, each mass spectrum was generated within half a second after tissue sampling. Ions were predominantly detected in the 600 to 900 m/z range, corresponding to 600 to 900 daltons (assuming all were in the expected single charge state).

An automated system was developed for real-time tissue recognition based on mass spectrometric profiles composed of preprocessing, multivariate statistical, and classification techniques. It can be divided into training and classification phases. In the training phase, the tissue classifier was trained to extract molecular ion patterns of histologically classified tissue types, which were subsequently used for prediction of unknown specimens. This required a knowledge-based database of raw spectra and corresponding histological classification for various tissue types. This database was constructed by the analysis of ex vivo tissue specimens removed during surgical resection of tumors. All collected data were stored in the Oracle database for easy use and access. During the training phase, a relevant subset of histologically assigned spectra for a particular organ system is selected for further analysis. All retrieved spectra of specimens were binned with a bin size of 0.1 dalton to minimize the influence of noise. Total intensity normalization was then applied to account for sample-to-sample variation in overall signal intensity, unrelated to molecular patterns. After this preprocessing, a set of correlated (redundant) molecular ion variables was transformed by a combination of multivariate statistical techniques into a lower dimensional set of nonredundant components for improved visualization, explorative analysis, and predictive modeling.

PCA was initially applied to map multidimensional REIMS data into an uncorrelated set of components capturing the majority of variation in the data set. Graphical representations of the first few components were used to explore the overall similarity/difference in molecular ion composition between tissue specimens. To ensure that the derived principal components explained systematic variation not attributable to noise, (3 × 3) bi–cross-validation was performed (15). The variance captured in bi–cross-validation was used as an indicator of model robustness (Rcv). Unless the total variation is dominated by the variation between various tissue types, the derived principal components are a suboptimal choice for extracting tissue-specific discriminating patterns (36, 37). LDA was therefore subsequently applied on the first 60 principal component scores to derive components with the enhanced capacity for discriminating between tissue types by taking into account the histological assignment of specimens. All classification models were validated by leave-one-patient-out cross-validation.

During the real-time recognition phase, the same preprocessing procedure was applied to an unclassified (test) specimen. After that, the discriminating molecular ion patterns of the test specimen were extracted by means of PCA-LDA projection matrix. The Mahalanobis distance was then computed between the extracted molecular ion pattern of test specimen and the molecular patterns of each tissue type. The test specimen was classified to the tissue type for which the Mahalanobis squared distance was minimal. However, if the distance of the unknown spectra point exceeded 5 SD in any dimension, the point was marked as an “outlier.”

All steps of the above analysis were based on computationally undemanding linear algebra operations and thus could be completed in real time. Univariate statistical analysis of the difference in the median values of individual peak intensities across tissue types was performed with the Kruskal-Wallis ANOVA-type test. The Bonferroni adjustment was used to account for multiple comparisons.

Supplementary Materials

www.sciencetranslationalmedicine.org/cgi/content/full/5/194/194ra93/DC1

Methods

Fig. S1. Schematic of intraoperative tissue identification using the iKnife.

Fig. S2. Statistical analysis of brain tissue data acquired in vivo and ex vivo.

Table S1. Patient cases where preoperative and postoperative histology results did not match.

Table S2. Histology-based diagnosis data for all patients recruited into the study (N = 393).

References and Notes

  1. Funding: The research was supported by the National Institute for Health Research (NIHR) Biomedical Research Centre based at Imperial College Healthcare National Health Service (NHS) Trust and Imperial College London. This work was also supported by the European Research Council under Starting Grant scheme (contract no. 210356), the Hungarian National Office for Research and Technology under Jedlik Ányos Grant scheme (JEDIONKO Grant), and TAMOP-4.2.1B-09/1/KMR-2010-0001. Author contributions: Z.T., J.B., and L.S.-S. designed the study. J.B. and L.S.-S. acquired the data. B.D. prepared histology and examined tissues for validation. J.B. analyzed the data. J.K. and L.J.M. were responsible for correlation of clinical data. M.R.L. and K.V. provided statistical support and interpreted the data. Z.T., J.K.N., L.D., and A.D. supervised the study and provided knowledge and support in the overall design and execution of the study. Z.T., J.B., L.J.M., K.V., R.M., and J.K. wrote and revised the manuscript. Competing interests: The views expressed are those of the authors and not necessarily those of the NHS, the NIHR, or the Department of Health. Z.T. serves as a senior consultant for MediMass. J.B. works as Head of IT at MediMass. The success of the REIMS/iKnife technology may increase the value of MediMass, but market introduction of the method is not anticipated within the next 3 years. A patent application disclosing the basic principles of REIMS-based tissue identification is currently pending. Data and materials availability: All spectroscopic data, including the reference database, used in the current study are available per request addressed to Z.T., upon agreement on confidential disclosure of the data to third parties. The data will be openly available three calendar years after the online publication date of the paper.
View Abstract

Navigate This Article