Rapid evaporative ionisation mass spectrometry of electrosurgical vapours for the identification of breast pathology: towards an intelligent knife for breast cancer surgery

Breast tissue samples were trimmed to size (3–10 mm²) and between 1 and 10 small cuts were made through the tissue using a modified monopolar blade electrosurgical pencil (Medres, Hungary) in either the pure cut setting (continuous RF wave) or fulgurate coag (pulsed RF wave) setting with a ForceTriad™ generator (Medtronic, Ireland). The power setting of the device varied between 10 and 30 W depending on the size and type of tissue. Aerosol produced as a result of electrosurgical activation was aspirated via the electrosurgical hand-piece [41] and transferred through a plastic tube to the mass spectrometer using a Venturi air jet pump. Surgical aerosol was co-aspirated with propan-2-ol (Sigma, MO, USA) (0.2 ml/minute) into the vacuum system of the Xevo G2-XS quadrupole time-of-flight mass spectrometer (product use is investigational) (Waters, UK). Aerosol particles and solvent droplets were de-clustered using a heated jet disruptor surface in the coarse vacuum regime of the instrument. Gaseous negative ions then entered the MS ion optics and were subjected to mass analysis. The remaining tissue was transferred to histology cassettes and sent to the pathology laboratory to be formalin fixed, paraffin embedded, sectioned and stained with haematoxylin and eosin (H&E). Subsequently, H&E-stained slides were examined by senior histopathologists to identify the tissue surrounding the sampling point and assign a tissue diagnosis (i.e. B1 = normal, B2 = benign, B3 = benign with uncertain malignant potential, B4 = suspicious, B5a = in-situ, or B5b = invasive tumour) according to the UK Royal College of Pathologists guidelines for non-operative diagnostic procedures and reporting [43].

To ensure only samples with adequate spectra and representative of true pathological change were used to build the tissue-type MS database, strict inclusion and exclusion criteria were determined (detailed in Additional file 2: Table S2); 40 specimen data files were excluded from a total of 399, leaving 359 for analysis of normal (B1 and B2) versus tumour (B5a and B5b). Raw mass spectrometric data were processed with Offline Model Builder (OMB-v29, Waters Research Centre, Hungary) with a bin of 0.01 m/z with background subtraction and lock mass correction (699.497 m/z). Spectra acquired from all sample burns were averaged to produce a single mass spectrum per sample. The mass spectra from all samples were imported into Matlab 2014a (Mathworks, MA, USA) and profile mode alignment [44] was applied prior to peak picking. The data were median fold change normalised and log transformed prior to multivariate statistical analysis. Principal component analysis (PCA) was used to identify trends in the data. Linear discriminant analysis (LDA) was used to identify spectral differences between cancer and normal tissue. Classification performance was recorded for each model with a leave-one-patient-out cross-validation scheme.

A sub-set of high-density (90–100%) cancer samples (n = 17) was compared to randomly selected normal breast tissue (n = 17) to determine the differences in peak intensities between cancerous and normal tissue. Total intensity normalization was performed prior to univariate analysis. The Mann-Whitney U test was performed, with Benjamini-Hochburg-Yekutieli false discovery rate correction accounting for multiple testing (p < 0.05). Isotope peaks were excluded. Thawed frozen normal (n = 2) and cancerous (n = 2) breast tissues were sampled in cut mode for 10–15 seconds to allow fragmentation of ions and collection of tandem MS (MS/MS) data. Tentative ion identification was performed by searching peak m/z values in the METLIN metabolite database and with LIPID MAPS® online tools [45, 46], which was refined using the MS/MS data.

The accuracy of the combined cut and coag ex-vivo model was tested by exporting the OMB statistical model data into purpose-built OMB recognition software (V29, Waters Research Centre, Hungary). The following parameters were set: mass range 600–1000 m/z, 0.1 bin, background subtraction on and lock mass correction to 699.497 m/z (phosphatidic acid (PA) (36:2)). REIMS analysis was performed in cut and coag modalities on new fresh and defrosted, normal and tumour breast tissues with both macro and microscopic histological agreement of tissue type. The validation spectra were pre-processed as previously described, then transformed to the linear discriminant space and classified to the closest class (i.e. normal or tumour) within the space using Mahalanobis squared distances. Recognition output was compared to histopathological results from H&E slides of the same tissue sample. For the correct classification of normal tissue, spectra from all sampling points within the specimen must have been registered as “normal” breast. Conversely, for tumour, at least one spectrum detected within a sample must have been considered positive for “tumour”. Figure 1a illustrates the ex-vivo recognition workflow. For additional validation, real-time ex-vivo tumour detection was performed on three mastectomy sections using electrosurgical dissection through normal breast into tumour tissue and back to normal breast tissue. The acquisition of REIMS spectra was synchronized with a video recording (GoPro, CA, USA) of the smoke capture as the tissue was continuously cut or coagulated with an electrosurgical blade. OMB recognition software was used to classify tissue content and this was compared to macroscopic tissue observations.

To determine if the ex-vivo method was applicable to the intraoperative environment we ran an intraoperative proof-of-principle study. A modified Xevo G2-XS mass spectrometer (Waters, UK) was installed in the operating theatre and a commercially available sterile (Surg-N-Vac or AccuVac, Covidien, UK) hand-piece was connected to the instrument. Aerosol produced as a result of electrosurgical tissue manipulation was continuously aspirated into the instrument throughout the operation. Video footage was simultaneously recorded, capturing all activities occurring at the operative scene (GoPro, CA, USA) in a time-synchronized manner with the acquisition of spectral data. Video recordings enabled retrospective orientation of spectral data with regard to three-dimensional margins. However, presently sample size is inadequate for interpretation of diagnostic accuracy. An optimised iKnife intraoperative workflow is displayed in Fig. 1b.

Average spectra obtained from 253 normal (B1 and B2) and 106 tumour (B5a and B5b) samples were used to create typical REIMS multispectral “fingerprints” for normal and cancerous breast tissue in both cut and coag electrosurgical modalities (Fig. 2). Spectral feature intensities were significantly dependent on the tissue type and electrosurgical setting used. Normal tissue demonstrated high intensity spectra in the phospholipid range (600–850 m/z) and the triglyceride range (850–1000 m/z) using cut mode and predominantly in the triglyceride range using coag mode. Conversely, tumour tissue demonstrated an increase in the phospholipid range and a decrease in the triglyceride range in both cut and coag modalities.

Statistical models were created for normal breast (B1 and B2) versus breast tumour (B5a and B5b) using cut, coag or a combination of the two (combined model). Data from 1158 sampling points from a total of 359 individual specimens; 253 normal (932 sampling points) and 106 tumour specimens (226 sampling points) from 113 patients were used to build the combined model (Table 1). The combined model was selected for recognition as both cut and coag modalities are frequently used interchangeably during breast cancer surgery. For the combined model, tumour class was distributed as 42 IDC, 8 ILC, 4 IMC and 2 DCIS (Table 1 and Additional file 1: Table S1). The mean (± standard deviation (SD)) age of patients in the combined group was 57.2 years (±13.48). The leave-one-patient-out cross-validation of the corresponding statistical model based on individual specimens (n = 359) resulted in overall accuracy of 94.4%, sensitivity of 93.4% and specificity of 94.9% (Fig. 3). For completeness, the cut model resulted in accuracy of 95.8%, sensitivity of 94.7% and specificity of 96.2% (Additional file 3: Figure S1), whilst the coag model resulted in overall accuracy of 94.7%, sensitivity of 93.9% and specificity of 95.0% (Additional file 4: Figure S2).

Univariate statistical analysis was used to identify significant peaks that differed between normal and cancerous tissue. After exclusion of isotopes, 24 significant peaks remained; 18 were higher in tumour tissue, with an average of 0.5 log₂ fold increase compared to normal and 6 peaks were lower in tumour, with an average of 5.7 log₂ fold decrease (Fig. 4 and Additional file 5: Table S3). The five most significant variables with higher cancer intensity are plotted in Fig. 5a, and the five variables with higher intensity in normal samples are plotted in Fig. 5b. MS/MS experiments were performed to identify the lipid species. Table 2 highlights the phospholipid species identified from the MS/MS spectra that were elevated in cancer compared to normal tissue (600–850 m/z range). All of the lipids identified in this range were glycerophospholipids with phosphatidylethanolamines (PEs) being the most commonly identified lipid species, followed by phosphatidylcholines (PCs), phosphatidic acids (PAs), phosphatidylserines (PSs) and phosphatidylglycerols (PGs) in that order. Table 3 summarises the triglyceride species identified from the MS/MS spectra of peaks that were significantly lower in cancer (850–1000 m/z).

New frozen and fresh breast specimens (n = 260) not previously used for ex-vivo model creation were subjected to iKnife analysis using both cut and coag modalities, employing an identical ex-vivo methodology combined with interpretation by bespoke iKnife recognition software (OMB V29). For the cut modality there were 79 normal and 48 tumour specimens and for the coag modality there were 82 normal and 51 tumour samples. Recognition software was concordant with final histopathological assessment in 249 out of 260 specimens, producing an overall model accuracy of 95.8% with a sensitivity of 90.9% and specificity of 98.8% (Fig. 6). Finally, the iKnife was used to perform a continuous dissection line through the tissue in cut and coag mode from normal, through tumour returning to normal tissue, in a case series of three whole mastectomy slices with large tumours (>3 cm). OMB recognition results demonstrated good overall classification (224/231 spectra, 97% accuracy) with video synchronised macroscopic tissue correlation (Fig. 7).

The iKnife was used during the entire surgical intervention in six case studies as a proof of principle of the intraoperative method. The mean time from electrosurgical activation to detection of online analysis was 1.80 seconds (SD ±0.40). High intensity spectra (total ion count (TIC) = e7-e8) were obtained in all surgical interventions both in cut and coag mode throughout the operation and were comparable in intensity (TIC = e6-e8) and overall visual morphology to the ex-vivo spectra (Fig. 8). Of the intraoperative spectra across six surgical interventions, 99.27% (n = 5422/5462) were interpretable by the ex-vivo model. Only 0.73% (n = 40/5462) of spectra were classified as outliers (greater than 2 SD) according to the ex-vivo classification model.

Due to the width of the electrosurgical blade (4 mm) the iKnife has relatively low resolution that may lead to dilution of tumour cellular content by normal cells, which may provoke a false positive result. REIMS is a destructive process, therefore it is impossible to be certain of the histology of the exact cells under analysis. The need to perform REIMS analysis on tissue prior to formalin fixation and microscopic assessment limited the number of cases of ductal carcinoma in situ (B5a) that were available for inclusion. There is uncertainty about how the iKnife will classify solid benign lesions. In a sub-set analysis of tumour (B5b) versus fibroadenoma (B2) (Additional file 6: Figure S3), good sensitivity (94.1%) but poorer specificity (87.3%) was observed, indicating that the spectral differences between these two groups are subtler than comparisons between either of these entities and normal tissue.