Towards intra-operative diagnosis of tumours during breast conserving surgery by selective-sampling Raman micro-spectroscopy

All breast tissue samples (from 60 patients) were provided by the Nottingham Health Science Biobank and were obtained during routine breast cancer surgery at the Nottingham University Hospitals National Health Service (NHS) Trust. Tissue sections were cut from blocks removed during surgery into 20 µm sections. The samples were kept frozen at −20 °C until used for Raman spectral measurements. Adjacent tissue sections (7 µm) were stained by haematoxylin and eosin (H&E) and the diagnosis was provided by two histopathology consultants. Only samples for which the diagnoses provided by the two histopathologists agreed were included in the study.

The Raman spectra were recorded using a custom-built Raman micro-spectrometers based on an inverted optical microscope (Eclipse-Ti Nikon) with an automated sample stage (H107 with Proscan II controller, Prior Scientific) and a 50 × / 0.55 objective (Leica), 785 nm laser (Starbright XM Torsana, power 100 mW at sample surface), 77 200 Oriel spectrograph equipped with a deep-depletion back-illuminated cooled CCD (DU401-A-BR-DD, Andor Technology). The acquisition time for the Raman measurements was 2 s per position. The spectrometer was calibrated using naphthalene and 1,4-bis(2-methylstyryl) benzene samples (both from Sigma-Aldrich, UK) to an accuracy of 0.5 cm⁻¹. The auto-fluorescence images of tissue sections were measured using a wide-field fluorescence imaging system integrated to the Raman microscope: 365 nm high-power LED (Thorlabs), CCD camera (01-QIClick-F-M-12 Mono, QImaging) and fluorescence filters (Semrock, DAPI-5060 C-NTE-ZERO). A 10 × / 0.25 objective (Leica) was used to obtain the fluorescent images with 500 ms integration time. A number of images were stitched into a single tissue image.

Raman spectral maps were recorded from tissue sections deposited on MgF₂ discs by raster scanning with 5 or 10 µm step sizes (total 10 000 spectra per sample). The following pre-processing procedures were applied to all Raman spectra: removal of cosmic-ray peaks, subtraction of the background Raman signal of the microscope objective and MgF₂ substrate, subtraction of a baseline based on a rubber band method (the rubber bands were chosen between the regions 750–850 cm⁻¹, 855–950 cm⁻¹, 951–1050 cm⁻¹, 1135–1254 cm⁻¹, 1488–1600 cm⁻¹, 1675–1750 cm⁻¹), and normalisation to zero mean and unit variance. For each spectral map a 3 × 3 moving average filter was applied and followed by a two-step k-means clustering, with k₁ = 6, k₂ = 10. After the Raman spectral measurements, the analysed tissue section was compared with the adjacent H&E-stained section and the tissue structures were identified by the histopathology consultants. The pseudo-colour spectral images obtained by k-means clustering were correlated with the tissue structures revealed by the H&E images. The centroid spectra of every cluster containing at least 10 measurement sites was included in the database and labelled using the following five classes: ductal carcinoma of non-specific type (DC NST), tumour-surrounding inflammatory stroma (Infl. Stroma), lobules and ducts, stroma, and fat. A maximum of 5 centroid spectra per class per sample were included in the database. For the classification model, a total number of 230 000 Raman spectra were recorded from 45 patients. After the k-means clustering, 627 centroid spectra were used in the classification model: 229 for DC-NST, 95 for tumour-surrounding inflammatory stroma, 51 for lobules and ducts, 151 for stroma, and 101 for fat.

Several data reduction methods (principal components followed by the multi-class linear discriminant analysis (LDA) rank reduction to obtain four features for each spectrum) and classification techniques (LDA, quadratic discriminant analysis, multinomial logistic regression (MNLR) were compared based on a 5-fold cross-validation algorithm. All classifiers were tuned to the regime providing 95% sensitivity on the training data. Hence, the classification model was set for 95% sensitivity level as first priority for DC-NST, and second priority for tumout-surrounding inflammatory stroma. The rest of classes were detected at their maximum level of probability without any prioritisation.

The Raman spectra from a selected region of the tissue were acquired at 5 or 10 µm steps. After pre-processing, all spectra were analysed by the two-step k-means clustering algorithm to identify the tissue regions. The diagnosis of each region in the pseudo-colour spectral image was obtained by applying the classification model to the centroid spectrum of each cluster.

For each tissue section, the auto-fluorescence intensity images corresponding to collagen (excitation 365 nm/emission 447 nm) were segmented using an unsupervised segmentation algorithm (O'Callaghan and Bull 2005) and Matlab code developed by Naidu (Naidu 2010). The edge pixels of the segments were trimmed off in order to avoid sampling points for Raman spectral measurement at the segment boundaries. To avoid the out-of-focus problem in Raman spectroscopy, the measurements were divided based on a 3 × 3 grid, and focusing was manually adjusted at each region. Segments smaller than 0.02 mm², normally found within the stroma, were ignored and were labelled as stroma without any Raman measurement. The following pre-processing was applied to the Raman spectra: removal of cosmic-ray peaks, subtraction of the background and baseline, normalisation to zero mean and unit variance, followed by de-noising based on an independent set of Raman spectra from five patients (40 000 spectra) using principle component analysis with 50 loadings (Marcsisin et al 2012, Kong et al 2013). The classification model was applied on each individual spectrum and the diagnosis of each segment was based on majority-voting. A segment was diagnosed as DC NST or tumout-surrounding inflammatory stroma if at least two spectra were classified as belonging to these classes. For spectra for which the classification uncertainty as measured by the Shannon entropy was 5% higher than the maximum value (i.e. log K, where K = 5 is the number of classes) were labelled as 'Unknown' (most unknown spectra were found to correspond to holes in the tissue).

The first step in the development of the spectral classification model for breast tumours was to establish a method to assign the measured Raman spectra to the correct tissue structures in the breast tissue samples. Figure 1 shows typical examples of breast tissue samples containing ductal carcinoma of no special type (DC NST) and normal tissue. By comparing the pseudo-colour images obtained by k-means clustering of the Raman spectra with the histopathology image obtained by H&E staining, the centroid spectra for each cluster were assigned a label corresponding to tumour (invasive carcinoma) or normal breast tissue (fibro-fatty tissue and parenchyma). The centroid Raman spectra show that significant differences can be detected between Raman spectra of tumours and normal tissue that originate from their different molecular compositions.

Zoom In Zoom Out Reset image size

Figure 2(a) presents the mean of the Raman spectra corresponding to ductal carcinoma no special type (DC NST), surrounding inflammatory stroma (Infl. Stroma), lobules and ducts, stroma and fat. The results show that the Raman spectra of DC have more intense spectral bands assigned to nucleic acids (788 cm⁻¹, 1098 cm⁻¹) compared to other tissue structures, while the Raman spectra of stroma are dominated by intense bands assigned to collagen fibres (851 and 950 cm⁻¹). These findings are consistent with the H&E images showing a higher number of cells with enlarged nuclei in the tumour regions (figure 1), and with the Raman spectra of other tumour types reported in the literature, such as basal cell carcinoma (Nijssen et al 2002, Larraona-Puy et al 2009, Kong et al 2013), bladder (de Jong et al 2006), oesophagus (Shetty et al 2006) or cervical cancers (Lyng et al 2007). The fat regions showed specific Raman bands characteristic to lipids (850 cm⁻¹, 1070 cm⁻¹, 1267 cm⁻¹, 1301 cm⁻¹, 1450 cm⁻¹, ~1660 cm⁻¹) (Tu 1982). The Raman spectra of tumour-surrounding inflammatory stroma were found to have lower intensity bands corresponding to collagen and higher intensity for the bands assigned to nucleic acids compared to the Raman spectra of normal stroma.

Zoom In Zoom Out Reset image size

Several spectral classification models were tested based on various data reduction and statistical classification methods, and were optimised for diagnosis of DC NST as first priority and tumour-surrounding inflammatory stroma as second priority. These two classes were grouped together in the first step of classification because of the clinical relevance and high similarities between the Raman spectra. The best performance for detection of tumours (highest specificity when sensitivity was set at 95%) and consistence between the training and validation datasets was obtained when 10 principal components of the raw spectra were further reduced to K − 1 (with K = 5 being the total number of tissue classes) canonical features using the rank-reduced multi-class linear discriminant analysis, then used with the multinomial logistic regression classifier. At a target sensitivity of 95%, the cross-validation results indicated 95.6% sensitivity and 96.2% specificity for classification of DC NST and the surrounding inflammatory stroma. The confusion matrix for the cross-validation is shown in the figure 2(b). The most misclassifications occurred between the 'lobules and ducts' class and DC. It is worth noting that the specificity was not affected by misclassification between DC-NST and tumour-surrounding inflammatory stroma. The receiver-operating characteristic (ROC) is presented in figure 2(c), and it shows consistency between the training and validation dataset, with the validation accuracy slightly less training accuracy, as this is expected for the cross-valuation. The specificity for DC diagnosis decreases from about 100% at 85% sensitivity to 82% at 100% sensitivity. The area under the ROC curves for the training set is 0.994, and validation set is 0.973.

Although raster-scanning requires long acquisition time that are impractical for intra-operative diagnosis, we first established the full capabilities of Raman microscopy to image and diagnose DC for independent tissue samples under this idealised setting. Figure 3 presents the diagnosis images for five independent breast tissue samples (four positive and one negative) when the target sensitivity of the classification models was set at 95%. The results show that the tumour regions were correctly diagnosed by the classification model, confirming the high sensitivity and specificity of Raman spectroscopy to detect molecular changes in tissue related to mammary invasive carcinoma. As the corresponding specificity of the classification model was ~96%, for each sample (60 clusters per sample) we expected, on average, 2.4 clusters corresponding to healthy tissue to be misclassified as DC (false positives). Such misclassifications can be observed for example in the diagnosis images in figure 3(b), where epithelium regions were mislabelled as ductal carcinoma or tumour-surrounding inflammatory stroma (blue arrows). In certain tissue samples, 100% specificity for tumour was observed, as shown in figure 3(e).

Zoom In Zoom Out Reset image size

To reduce the data acquisition time associated with the raster-scanning, tissue auto-fluorescence imaging can be used to determine the main spatial features a tissue sample, and this information is then used to select and prioritise the sampling points for Raman spectroscopy (Kong et al 2013). To evaluate this sampling strategy, multimodal spectral histopathology (MSH) was compared with raster scanning Raman microscopy images using 2 × 2 mm² tissue regions containing DC and other tissue structures. Figure 4(a) shows that the auto-fluorescence imaging has high sensitivity for detection of tumours (blue arrows) as tumours can be well delineated. However, the specificity of tissue auto-fluorescence is low, as other tissue structures with low collagen content, such as epithelium and fat, also elicit similar auto-fluorescence intensity to DC. Rather than using the auto-fluorescence intensity for diagnosis, the unsupervised image segmentation algorithm (O'Callaghan and Bull 2005, Naidu 2010) was used to divide the auto-fluorescence image into segments and determine the structural pattern of the tissue sample as shown in figure 4(b). For every segment, a space-filling algorithm was used to generate 5 sampling locations for the Raman spectral measurements. Figure 4(c) shows that after the segmentation of the 2 × 2 mm² tissue region into 117 segments, only 585 Raman spectral measurements were required to obtain a correct diagnosis of the tissue sample. Figure 4(d) shows that the diagnosis images obtained by raster scanning Raman spectroscopy and MSH were in agreement with the histopathology diagnosis based on the H&E-stained adjacent section (figure 4(e)). However, in this case, MSH provided a dramatic decrease in the acquisition time as it required 68-fold fewer Raman spectra compared to raster scanning, while providing a similar diagnosis accuracy for DC. A slight drop in diagnosis accuracy was observed in the MSH image compared to raster scanning, which was attributed to the fact that in MSH the classification model was applied to individual Raman spectra, which had lower signal-to-noise ratio compared to the centroid spectra used for the classification of the raster-scanning images.

Zoom In Zoom Out Reset image size

The selective-sampling was then implemented to provide diagnosis of eight new independent breast samples. Typical examples of diagnosis of ductal carcinoma are presented in figure 5. While the number of Raman spectra required for the diagnosis of the tissue samples depended on the size and structural complexity of the samples, the results show that for typical 5 × 5 mm² samples, accurate diagnosis by MSH can be obtained with ~500 Raman spectra. For the target 95% sensitivity level, the multimodal spectral histopathology successfully detected the tumours in all analysed tissue samples.

Zoom In Zoom Out Reset image size

For our laboratory-based instrument, the acquisition time for the auto-fluorescence images was 4 min and integration time for each Raman spectrum was 2 s. Using these estimates, the diagnosis time for MSH was 20 min, which is shorter than the typical diagnosis time by frozen section histopathology. However, as the MSH diagnosis is no longer limited by the time-consuming tissue sectioning and staining, the development of an optimised prototype instruments using optimised high speed Raman microscopes (e.g. integration time per spectrum of 0.2–0.5 s/spectrum) could allow intra-operative assessment of larger breast resections (~2 × 2 cm²) within 10–15 minutes. The key advantage of MSH over other techniques is that the diagnosis is objective and can be obtained without any tissue preparation (sectioning and staining).