Resolving Interobserver Discrepancies in Lung Cancer Diagnoses by Spectral Histopathology

In this section, the basic background of CIRECA_SHP and the underlying spectral technologies will be introduced. This review begins with a short discussion of FTIR spectroscopy, which is a common analytical technique in which molecular compounds are identified and quantified by a physical optical property, namely their infrared absorption spectra. These spectra result from the interaction of midinfrared photons (2.5–10-μm wavelength or 4000–1000-cm⁻¹ wave number) with quantized vibrational energy levels of the molecular compounds of the sample.¹¹  The resulting infrared absorption spectra are characteristic signatures of these compounds; see Figure 1, A. Spectral libraries of thousands of chemical and biochemical compounds are commercially available and accessible, and are used in a wide variety of analytical applications ranging from structural studies of molecular species to identification of synthetic products and quality control in the food and chemical industries.¹²  More recently, FTIR spectroscopy has found applications in the detection and identification of manufacturers of counterfeit drugs and threats to public safety (such as rapid detection of anthrax).¹³  It also has been used to detect art forgeries and general contraband. Fourier transform infrared spectra can be collected from a few milligrams of sample in a few seconds' acquisition time. The first commercial infrared spectrometers have been available since the 1950s, and FTIR spectrometers since the mid-1970s.

In infrared microscopy, the smallest sampling area (pixel) from which a spectrum can be collected is determined by the Rayleigh criterion (also known as the diffraction limit), which determines a pixel size of about 10 μm on edge, depending on the wavelength of light and the numerical aperture of the microscope objective. Given the size of a small human cell (20 μm in diameter), it follows that about 4 infrared pixels are required to image such a cell. Most commercial infrared microscopes collect spectra from somewhat smaller pixels, 5 to 6 μm on edge, depending on optical design criteria. With visible light (with a wavelength of 500 nm) one can image objects that are approximately 10-fold smaller than is possible with infrared light.

An infrared pixel spectrum represents a snapshot of the chemical or biochemical composition of that pixel. By moving the sample, via a computer-interfaced microscope stage through the focus of the microscope, spectral data sets can be constructed¹⁴  that derived their imaging contrast by spatial variations of the infrared spectra, and therefore by the sample chemical composition. Visualization of these variations will be discussed later.

If the samples are solids, the thickness of the sample, which determines the amplitude of the observed spectrum, should be about 10 μm or less because the absorption cross section, that is, the efficiency with which a molecular compound absorbs infrared radiation, is so large that thicker samples absorb too much of the incident radiation. At a density of about unity, such a sample volume (10³ μm³) corresponds to a sample mass of about 1 ng, or 10⁻⁹ g. A midinfrared spectrum collected from such a sample volume may consist of 1500 absorption intensity data points, typically spaced at 2-cm⁻¹ intervals in the 1000- to 4000-cm⁻¹ spectral range (10- to 2.5-μm wavelength range). This is shown in Figure 1, A, which depicts typical pixel spectra for different tissue types. All 3 spectra show characteristic peaks due to peptide linkages found in all proteins (the protein amide I and amide II bands); these peaks always are prominent in spectra of cells and tissue because proteins, by dry mass, are the most abundant cellular components. Trace a of Figure 1, A, also shows prominent carbohydrate peaks in a tissue type that is known to store glycogen. Trace b of Figure 1, A, is from connective tissue and shows characteristic peaks due to collagen (see marks); these peaks are particularly obvious in second derivative representation (see below). Finally, trace c of Figure 1, A, shows the spectrum of B lymphocytes; these spectra exhibit pronounced nucleic acid (DNA and RNA) features.¹⁵ 

Modern commercial FTIR microscopes are equipped with array detectors that allow the simultaneous acquisition of up to 16 384 (128 × 128) pixel spectra, each from an area of 5.5 × 5.5 μm, for a total field of view of 0.7 × 0.7 mm. A typical data acquisition time for such an infrared spectral image data set, consisting of 16 384 individual spectra with each comprising 1500 intensity data points, is about 7 minutes.

Since the early 2000s, FTIR microscopes have been used to collect spatially resolved image data sets of human cells and tissue.¹⁶  As pointed out before, such data sets may contain thousands of pixel spectra, with each spectrum consisting of about 1500 data points depending on instrument software design. These data sets were originally analyzed, or visualized, by plotting the intensity of a given peak in the infrared spectrum as a grayscale or color-coded feature versus the position at which each spectrum was collected, yielding univariate maps or images of the concentration of one component at one wavelength. Subsequently, multivariate methods were applied to the analysis of imaging data sets.^17–19  In one of these methods (known as unsupervised hierarchical cluster analysis [HCA])²⁰  the entire spectrum is used, and a correlation of each spectrum with all other spectra in the data set is performed. Subsequently, the values of the correlation coefficients may be displayed at the pixel coordinates, again as grayscale or pseudocolor images, to reveal similarity of the pixel spectra. This similarity, in turn, is related to variations in biochemical composition in the sample. The term unsupervised introduced above indicates that only spectral data and their similarity are being used to construct images and that the underlying medical conditions are unknown. Examples of such unsupervised, HCA-derived images are shown in Figure 2, A through C.

Figure 2, B, represents an HCA-based image of a TMA sample containing connective tissue, small cell lung carcinoma, and necrotic tissue at the 5-cluster level. The cluster level represents the ability of the clustering process to differentiate among tissue types; higher cluster levels correspond to increased discrimination of spectral differences. Figure 2, A, represents the corresponding H&E-stained image of the same tissue section. In Figure 2, B, the areas indicated in light blue represent spectrally homogeneous regions that can be associated with necrosis by comparison with the H&E image. Cancerous regions are shown in dark blue and connective tissue in red and green colors, the differences being due to density of immune cell infiltration in the connective tissue. Thus, the HCA-based results depict, in a completely unsupervised approach, tissue compositional variations that exhibit a remarkable similarity to the regions and boundaries found in the H&E-stained images.¹⁶  (It should be noted that the color assignment of the clusters in unsupervised methods is arbitrary.)

Figure 2, C, represents an HCA-based image at the 9-cluster level. At the increased cluster level, as compared with Figure 2, B, more tissue types are distinguished, resulting, for example, in the detection of blood vessel walls (gray, V-shaped area at the 2-o'clock position in Figure 2, C) and a more refined description of the immune cell infiltration in the connective tissue.

Interestingly, the necrotic tissue shown in Figure 2, B and C, exhibits an infrared spectrum in the amide I spectral region that is very different from, for example, cancerous tissue; this is shown in Figure 1, B. This difference results from the presence of partially unfolded and presumably precipitated proteins that show an additional amide I component (see Figure 1, B) at circa 1638 cm⁻¹. This peak, associated with unfolded proteins, is particularly obvious in the second derivative spectra in Figure 1, B. The use of second derivative spectra will be discussed below in more detail.

The methodology of using infrared spectral images and their correlation to classical H&E images of tissue has been referred to as SHP. The colocalization of features observed in the H&E-stained and the infrared spectral images confirms the well-known fact that the components of H&E stain, or any IHC stain, bind to different chemical moieties and therefore create an image contrast based partly on chemical characteristics of tissue. In SHP, however, this image contrast is created solely by chemical heterogeneity that is observed via an inherent, objective, reproducible, and instrument-based signature.

Spectral histopathology has been applied to a number of medical problems, mostly for the detection and characterization of cancer in various organs, in particular brain,^21,22  breast,²³  cervix,^24,25  colon,^26,27  lung,²⁸  prostate,^29,30  and a few others. These studies confirm the overall accuracy of SHP, although many of the reported results were obtained on small data sets, and often more in the format of pilot studies.

CIRECA_SHP, as indicated above, is a further development of the principles of SHP. This development includes pixel-level correlation and traceability of infrared, H&E-stained, and IHC images, as well as the development of supervised and unsupervised analytical methods for such data sets.

In supervised methods, classifiers are trained to recognize spectral patterns established in the training phase of the algorithm and render a diagnosis or classification result based on the similarity of unknown spectra to the training data. To this end, spectral signatures are collected from regions of well-defined histopathology. This first step of the supervised analysis is henceforth referred to as annotation. In CIRECA_SHP, this step is carried out via a software package, CIRECA_Annotate, that allows a pathologist or researcher to display H&E (Figure 3, A), IHC, and/or HCA spectral images that are automatically scaled, registered, and optionally overlaid. Next, regions of well-characterized histopathologic features are selected in the H&E, IHC, or HCA images by drawing outlines via a mouse or a stylus, as shown in Figure 3, B, and attaching a particular tissue or disease label. These labels and associated tissue codes are available from a pull-down menu. The outlined regions in Figure 3, B, are annotated as ADC or adjacent normal tissue. Figure 3, C, shows these areas as highlighted purple and orange overlay colors. An optional overlay of the selected regions and the HCA-based images may be used to verify the spectral homogeneity of the selected areas.

The spectra annotated in the step previously described were subsequently extracted and stored in a database that contained more than 2 million pixel spectra from different patients and from regions with different tissue labels that indicate tissue types and disease states. These pixel spectra were used to train an algorithm to classify or analyze unknown spectral data.

These algorithms represent supervised multivariate methods of analysis, where the term supervised implies that annotation by a pathologist was used to establish a correlation between a tissue type or disease state and its characteristic spectra. There are several well-established methods to perform such classifications, such as artificial neural nets, support vector machines (SVMs), random forests, and others.^11,20  They all are based on a learning process to recognize specific spectral signatures in the training phase and identify these features within a data set of unknown composition. For the work reported here, SVMs were used, although the other methods, in general, give similar classification accuracies if published rules and precautions on the statistical procedures for large data sets are followed.³¹  Details of the data analysis methodology and statistical approaches of the previous studies have been reported.⁷ 

Support vector machines represent a classifier architecture that can be visualized relatively easily. The spectra in the training set are represented as points in N-dimensional space, where N is the number of features (data points) in the spectra. For example, if each spectrum consisted of 3 measured intensity values, each spectrum would be represented as a point in 3-dimensional space. A 2-dimensional plane would then be constructed to separate different classes of spectra. As pointed out earlier, each spectrum discussed here contains 500 data points; thus, a 499 (ie, N − 1)–dimensional hyperplane was used to separate the different classes in the training set (for example, cancer or NOT cancer). The term hyperplane is used here to indicate the morphology of the separator in N-dimensional space. Unknown spectra were subsequently classified by where they fell with respect to the hyperplane. In addition, the distance to the hyperplane indicates the confidence of classification: the farther a point from the plane, the higher the confidence of a correct classification (see also the discussion of Platt SVM below).

The testing of classifiers must be carried out on data sets that were not used in the training phase; otherwise, overfitting of the data will occur. Therefore, the spectral data set for testing the classifier, in principle, would not need to be annotated. However, all spots were annotated in this study, even if they were in the test set, to allow assessment of the correctness of the prediction algorithm when compared with the annotation image.

The analyses in the previously published studies^6,7  and the present study were carried out at the pixel and the patient level. For the former, only annotated pixel spectra from each patient were used in the analysis. The annotated spectra were split into training and test sets. The test set was not used for algorithm training and refinement, as pointed out above, although the true pathologic results were known. Once classifier training was complete, the test set was analyzed, and sensitivity and specificity of the predicted results were compared against the true pathologic diagnosis. The pixel-level tests were also carried out using a leave-one-out cross-validation (LOOCV) in which all patients but one were used for training a classifier that was subsequently applied to the left-out patient.

This pixel-based analysis is most revealing because a single-pixel spectrum represents the highest level of image decomposition, and thereby a measurement of highest spectral purity. From a statistical viewpoint, pixel-based analysis is the most traceable approach because for each pixel spectrum used in both the training and testing process, the true outcome is known. However, one should keep in mind that the inherent heterogeneity of pixel-level data restricts the accuracy of pixel-level classification. This is because 2 adjacent pixels may be from the nucleus or the cytoplasm of a given cell, and therefore may exhibit quite different spectral signatures, although they belong to and were annotated to the same tissue or disease type.

In real-life pathology, the diagnosis needs to be made on a patient basis. Thus, classification results were obtained on the patient level as well. Here, the majority of pixel spectra in annotated regions must conform to the pathologist's diagnosis, but a small number of misclassified pixels, particularly in areas adjacent to cracks and voids in the tissue, are allowed. Misclassifications occur mostly because of low spectral quality in areas of poor cellularity or tissue folding. Patient-based analysis was carried out using LOOCV.

For the CIRECA_SHP methodology, detailed work instructions for sample and substrate quality control and instrumental performance and environmental conditions were established. Detailed methods of data acquisition and data processing will be presented in the Sample Selection and Methodology section below. All spectral processing was carried out using software developed at CIRECA, LLC, for the automatized preprocessing, HCA image generation, and annotation using either H&E or IHC images. MATLAB (MathWorks, Natick, Massachusetts) scripts have been developed, characterized, and tested for the extraction of annotated pixel spectra into data management software. Other MATLAB scripts are available for creating randomized training/test data sets, and for subsequent training of SVM classifiers, using optimized parameters.⁷  The output of the SVMs has been standardized to produce images of the annotated regions, images of the predictions in the annotated images, and images based on whole-spot analysis, that is, analysis of the entire spectral data set, including annotated and not-annotated regions.

Before discussing details of sample selection and methodology, we wish to reiterate that SHP in general and CIRECA_SHP in particular work on completely different principles than other recent computer-based pathology approaches, for example whole-slide imaging. In whole-slide imaging, an algorithm is trained to recognize morphologic features or staining patterns. The resulting classifier then identifies occurrences of these features in other regions of the slide, or on other slides. Spectral histopathology, in contrast, is based on recognizing spectral fingerprint signatures that are related to biochemical composition rather than morphologic features. The spectral features are inherent signatures of molecular composition and can be collected with extremely high reproducibility. Whereas morphologic features of the same disease or tissue type may vary from patient to patient, the biochemical signatures are maintained to a greater degree.

Furthermore, spectral data are more suitable for analysis by computer-based methods of multivariate analysis, because the data naturally are represented in vectorized form, that is, as spectral vectors of 500 intensity versus wave number (or wavelength) data points. In whole-slide imaging, morphologic features need to be vectorized by establishing certain metrics (such as nuclear to cytoplasmic ratios, cell size, etc) that are inherently less quantitative than a biochemical signature captured via infrared spectra.

The study reported below was carried out under an institutional review board protocol from COH Medical Center. The total sample set consisted of 450 patients (15 TMA blocks with 30 patients each), with 1 normal and 2 cancer tissue cores for each patient, for a total of 1350 patient samples. It was decided at this point to reduce the number of samples included in this study in order to carry out this study in a reasonable time span, to limit the amount of raw data created, and to restrict the time and cost associated with the annotation process, in particular because previous publications^7,31  established that the minimum number of patients required in each class for a given classification confidence level was much lower than the number of patients available in this data set. Thus, only 1 of the 2 cancer cores was used in this study, and patients were selected where unambiguous patient metadata were available.

The use of TMAs for this study may have biased the accuracy of SHP due to the fact that only unambiguous areas from a resected tumor may have been selected for inclusion in the TMA. Although this is certainly a possibility in a commercial TMA, it is less likely in a TMA that was assembled for a research study. Furthermore, selecting randomly 1 of the 2 cancer cores further reduces the chance for biasing the results because of the inherent heterogeneity between the cores. Furthermore, literature^32,33  on the subject of localized versus global sampling, as exemplified by a 1- to 2-mm² TMA core as compared with an entire resection, indicates that TMAs, in general, reproduce the global diagnoses quite well. In addition, we have examined a subset of patients where both cancer cores were analyzed and found good agreement between the SHP predictions. Thus, 328 samples were selected based on the COH core histology report and visual microscopic inspection of tissue section quality. Because the tissue sections were unstained at this point, the main criterion for inclusion in this study was good tissue quality in the cancer samples, that is, good coverage and few voids, cracks, and folds (normal lung tissue naturally is very sparse). Of the 328 tissue samples included in this study, 77 had been diagnosed by COH core histology as normal samples from cancer patients, 162 had been diagnosed as ADC, 75 as SqCC, and 14 as AdSqCC. The composition of the data set used in this study is summarized in Table 1.

From each of the 15 TMAs, four 5-μm-thick sections were used in this study. The first section (A001) was mounted on standard glass microscope slides, deparaffinized, and stained at COH using the TTF-1 antibody marker (Leica Biosystems, Nussloch, Germany) and immunoperoxidase counterstaining to confirm the present of ADC. The second section (A002) was used for infrared imaging data acquisition. This section was mounted on a special, infrared-compatible slide, known as a low-emissivity slide. Glass cannot be used as a substrate material for midinfrared spectral studies because it totally absorbs all midinfrared wavelengths. The low-emissivity slides are transparent to visible light but reflective in the infrared spectral region, and can be used for infrared data acquisition in reflection mode and, after subsequent staining, for visual image collection in transmission. Tissue sections were deparaffinized following standard protocols. Subsequent to infrared data acquisition, the tissue spots of section A002 were H&E stained at UMP, coverslipped, and imaged in house at ×20 magnification (see below). This procedure ascertains that the infrared and visual images were collected from exactly the same slide, which helps in image registration and annotation. The use of low-emissivity slides has come under criticism in the literature,³⁴  but we have shown that they can be used without problems if certain precautions are followed.³⁵  The third section (A003) was mounted on glass slides and stained with a p40 antibody (Biocare Medical, Pacheco, California) for confirmation of SqCC. Section A004 was stained for programmed death ligand-1 (PD-L1), but the results obtained will not be discussed here.

All infrared spectral data sets used in this study were acquired using an Agilent FTIR microspectrometer (Model 620/670, Agilent, Santa Clara, California) equipped with a 128 × 128-element HgCdTe focal plane array detector in reflection mode, which, when used with low-emissivity slides, will result in the observation of transmission/reflection (or “transflection”) spectral data.³⁶  Four individual tiles, each measuring 0.7 mm on edge, were stitched together to form one image 1.4 mm on edge containing 65 536 individual pixel spectra. Visual images of H&E- or IHC-stained tissue sections were collected at ×20 magnification using an Olympus BX51 microscope equipped with a high-resolution digital camera (Model QICAM Fast 1394, Qimaging, Burnaby, British Columbia, Canada) and a Prior Scientific, Inc (Rockland, Massachusetts), digitally controlled microscope stage and operated under Surveyor Software with Turboscan (Objective Imaging, Ltd, Cambridge, United Kingdom). All visual images were collected at ×20 magnification and were scaled to match the size of the FTIR images in the figures in this paper. Thus, the scale of the images shown can be estimated from the fact that the diameter of each tissue spot was circa 1.2 mm. Both the Agilent and Olympus microscopes are located in a humidity- and temperature-controlled laboratory at CIRECA.

All spectral data sets were preprocessed using CIRECA_SPP, version R852, using second-derivative spectra and resonance Mie scattering compensation by phase correction.³⁷  The advantages of converting absorption spectra to their second-derivative spectra (see Figure 1, B) have been previously documented in the literature,³⁸  and include detection of protein contributions that are more difficult to observe in the absorption spectra because of the broadness of their spectral features. The preprocessing software also produced HCA-based SHP images that were used, along with the H&E images, for the annotation described earlier. All tissue spots were annotated at UMP by a board-certified pathologist, using CIRECA_Annotate, version R930. The annotation identified regions of the tissue core sections in which the pathologist identified a given tissue type or disease state with high confidence. The main tissue categories were normal tissue, normal tissue proximal to cancer (also referred to as cancer-adjacent normal tissue or normal tissue with malignancy-associated changes [MAC]), cancer, and benign lesions. Each of the major categories had associated tissue types, such as dense fibroconnective tissue with or without inflammatory cell infiltration or fibroblast (in the case of normal tissue) or ADC and SqCC cancer types. The cancer types had further subcategories, such as lepidic, solid, papillary, etc. Spectra from regions identified and labeled by the pathologist, in turn, were extracted and used for further analysis. The total number of annotated regions and the number of extracted spectra are summarized in Table 2.

Extraction of annotated data was performed by a MATLAB script that links all annotated spectra and the diagnostic code used by the pathologist to describe the tissue type or disease state. Based on these codes, spectra are fed into searchable database management software. From this database, data sets were created, based on random selection, to establish training and test sets for the classification algorithms. Details of these steps have been previously published.^6,7 

All patient tissue samples were imaged at CIRECA, LLC, and annotated at UMP. There was a total of 51 cases (15.5%) where the histology report from COH and the UMP diagnosis differed. This may be because of a sampling issue (diagnosis at COH was based on the evaluation of the entire tumor whereas the pathologic annotation at UMP was based on an area of circa 1.1 mm², the average area of the circular tissue core section). Furthermore, there may be significant differences in the percentage of cancerous areas of a core section over the length of the core. The training of the classifiers was performed based only on spectra extracted by the pathology annotation.

The SVM classifier used a 2-level decision tree that classifies pixel spectra based on the variance between spectral classes. Similar decision tree structures have been used for classification schemes²⁸  because the differences between spectral classes to be distinguished may vary considerably. The decision tree used in this study is shown in Figure 4. The first-level algorithm distinguished normal tissue from not normal, that is, cancerous tissue. The next-level algorithm was trained to distinguish the main tissue types that were the targets of this study: distal normal, proximal normal (cancer-adjacent normal tissue with malignancy-associated changes), and macrophages on one hand, and SqCC, necrotic SqCC, ADC, and necrotic ADC on the other. A third-level algorithm was trained to classify keratinizing and nonkeratinizing SqCC as well as the subtypes of ADC (lepidic, acinar, solid, papillary, and micropapillary). Results of this study will be reported at a later date.

The SVM algorithms used here permitted prediction of confidence levels (known as the Platt confidence level; see above) of the classification decision. These confidence levels were based on the distance of a pixel spectrum from the hyperplane separating tissue classes, and allowed pixel spectra to be eliminated from the analysis if the confidence level was below a preset threshold.

The SVM classifiers reported here differed from those resulting from the previous study⁶  mainly because of different instrument platforms used and the absence of one cancer class (small cell lung carcinoma) in the data set. Although we found that the choice of instrumental platforms used (Agilent versus Perkin-Elmer instruments) influence the spectral results only minimally, we decided to train instrument-specific classifiers for the COH study, in particular, because it aimed to achieve subclassification of ADC, and it was felt that such fine classification was only possible with an instrument-specific method. However, during this study, the causes of very small, instrument-specific spectral differences were determined, and methods to compensate for such differences were established.³⁹  Because the COH TMAs contained only normal, ADC, SqCC, and AdSqCC tissue sections, the COH classifier reported here does not include small cell lung carcinoma.

The pixel-based analysis was carried out using a Platt SVM (see above) using a training set–test set split of 80% to 20% or in LOOCV mode. The sensitivities and specificities of both methods were the same within statistical error. The results shown in Table 3 are based on LOOCV results. In the confusion matrices in Table 3, the labels in the top row refer to the SHP prediction, whereas the labels in the leftmost column refer to the gold standard, annotation by pathology.

For Table 3, all pixel spectra were accumulated into 2 classes, normal and cancer, and necrotic annotations were not included. The sensitivity and specificity of this test were 96.3% and 97.1%, respectively. These results are very much in line with the accuracy for cancer/noncancer discrimination achieved in prior studies.^1,6,40 

For the distinction of ADC from SqCC, only spectra annotated by the pathologist as either ADC or SqCC were used, again either in training/test or LOOCV mode. We found very good accuracy for this discrimination, as shown in Table 4. The data sets here were balanced⁷  to exclude any bias based on the size of the data sets. The SqCC classification accuracy was 94.2% for pixels annotated as SqCC, and the ADC classification accuracy was 90.2% for pixels annotated as ADC.

The accuracy of the pixel-level results is somewhat lower than that of the patient-level results (see Tables 3 and 5) for reasons pointed out above and reiterated here. Pixel-based analysis represents the highest level of deaggregation, or the highest level of inhomogeneity, of the spectral data. This statement implies that 2 adjacent pixels may be somewhat different if they result from different structures in the same cell, for example, nucleus and cytoplasm. Therefore, such 2 pixel spectra may be classified differently. However, in the annotation process these pixels are aggregated into one class, although they may differ spectrally. Thus, the pixel-level results inherently take into account tissue heterogeneity, and therefore exhibit somewhat lower prediction accuracy.

Although the pixel-based results reported in the previous section are highly encouraging in terms of the sensitivity and specificity achieved, the patient-based results are clinically more significant. These results were obtained by LOOCV, in which all patients but one were used for training a classifier that subsequently was applied to the left-out patient.

For patient-level results, the annotated pixels were aggregated by tissue type label. If the number of predicted cancer pixels was greater than 400 (about 100 cells) in the annotated regions, and if the cancer type was correctly predicted, then the sample was considered a true-positive result. If fewer than 400 cancer pixels were predicted and the annotation did not detect cancer, the result was considered true-negative (normal). Using these criteria, the sensitivity of detection of cancer (both ADC and SqCC) was 99.6% and the specificity 97.4% (see Table 5) for a total of 327 samples. One sample was omitted from the analysis because it contained an annotation category (reserve cells) that occurred only once in the entire data set.

As pointed out before, the accuracy of the patient-based results is expected to be better than that of pixel-level analysis because of aggregation of pixels into regions the size of which ignores individual misclassified pixels. The accuracy of the classification reported in Table 5 is excellent and suggests that an automatic, SHP-based system could be used clinically for triaging or initial screening of samples that need no further scrutiny by classical histopathology or IHC. Such a system also could be used to nondestructively establish whether or not a sample for a clinical study, that is, on a TMA, conforms to the original histology diagnosis.

The patient-based analysis has the additional advantage that visual, patient-based images of the results can be presented. Such images may be restricted to annotated regions only, or may include the entire tissue sample. Results from both these methods are reported here. The nomenclature used for these images in the following discussion is as follows: T0 images depict UMP-annotated regions without SHP-based classification and constitute the true gold standard. The annotated regions are displayed in the same color codes used throughout this report (see Figure 4). T1 images depict the LOOCV-based predictions for the annotated regions, whereas T2 images depict the results for the entire sample, again in the same color code. In the following discussion (in particular, in the figure captions), tissue sections are identified as CHLCxx_AABB, where CHLC indicates COH lung cancer, xx the TMA identifier (01–15), AA the row, and BB the column number of the spot in the TMA.

Furthermore, the SVM classification images produced by CIRECA_SHP permit depiction of the Platt confidence levels (see above): white regions in the prediction images indicate regions of low prediction confidence. These images are referred to as T2P70 images, where T2 denotes whole-spot images and P70 the Platt prediction confidence limit.

We start the discussion of patient-based LOOCV-derived images with results indicating the power of CIRECA_SHP to discriminate between truly normal tissue and normal tissue in close proximity to a cancerous lesion. It should be understood that the “truly normal” tissue samples came from patients who were lung cancer patients as well. However, these biopsies were collected from sites far removed from the cancer and may be considered cancer free.

CIRECA_SHP readily discriminated between normal tissue and normal tissue in close vicinity of a tumor. In order to achieve this discrimination, normal tissue types were annotated to be either cancer-adjacent (normal tissue with MAC [normal/MAC]) or distal normal tissue, and the classifier was trained to distinguish these 2 normal tissue main types. This is demonstrated in Figure 5. Figure 5, A through D, depicts images of the H&E-stained normal tissue (Figure 5, A), the selected annotation regions (the T0 image, Figure 5, B), the whole-spot prediction (T2 image, Figure 5, C), and the prediction confidence image (T2P70, Figure 5, D) of a tissue sample diagnosed as normal lung tissue by both COH histology and UMP. This distal normal tissue is shown in red in Figure 5, B through D. However, a few pixels in Figure 5, C, seem to indicate normal/MAC tissue, but the prediction confidence image (Figure 5, D) indicates low probability for the normal/MAC tissue. Thus, the overall CIRECA_SHP classification is normal for this sample. In contrast, Figure 5, E through H, depicts a tissue sample in which UMP indicated small ADC areas, shown in purple in the T0 image (Figure 5, F). Figure 5, G, shows the prediction (T1) image in the annotated regions, and there is, indeed, excellent agreement between the prediction image and the true image. The whole-spot prediction image (Figure 5, H) shows the normal tissue as normal/MAC in yellow, in agreement with the annotation that indicated the presence of cancerous areas near the normal tissue. In addition, the selected cancerous areas were correctly predicted. This example demonstrated that truly normal tissue was distinguished from normal/MAC tissue with high accuracy, and that the presence of small cancerous lesions can be detected by the changes in adjacent normal tissue. This last aspect may again be of high clinical significance, because CIRECA_SHP indicates the proximity of cancerous cells in a tissue sample that is considered free of cancer cells.

We now turn to the discussion of the ability of CIRECA_SHP to distinguish between main cancer classes, in this particular case between ADC and SqCC. In the case of poorly differentiated or undifferentiated non–small cell lung cancers, this distinction can be very difficult by methods of classical histopathology,⁴¹  but accurate diagnosis is highly important because of adverse side effects of therapeutic agents in the case of incorrect diagnosis.²  The results shown in Table 6 are based on a classifier that used the raw UMP annotation only; that is, these results did not take into account any subsequent adjustments to the UMP diagnoses because of the availability of IHC results, which affected a few tissue samples. Thus, the results presented in Table 6 may be considered the worst-case scenario for SHP accuracy, and some of the CIRECA_SHP misclassifications were later corrected by IHC. A total of 240 samples (of 251 cancer samples; see Table 1) were included in this effort; 11 samples were rejected because of absence of cancerous cells.

Figure 6 presents examples of results obtained for ADC/SqCC classification where COH histology, UMP, IHC, and CIRECA_SHP all agreed, and may be considered a baseline for all methodologies used in this study. The results presented in Figure 6 are representative of the majority of cancer samples in this study, as indicated by the total ADC/SqCC discrimination accuracy reported in Table 6.

Figure 6, A and B, shows H&E and TTF-1 images, respectively, of a tissue sample diagnosed as ADC. Figure 6, C, depicts the T0 annotation (true) image, and Figure 6, D, the corresponding whole-spot prediction (T2) image. There is good agreement between the cancerous areas annotated in Figure 6, C, and the prediction of cancerous areas in Figure 6, D, except that CIRECA_SHP classified more of the tissue surrounding the annotated cancerous regions as cancerous as well. This may be because of the lower spatial discrimination of SHP or the detection of cancer signatures in the tissue directly adjacent to cancerous cells. Figure 6, E through H, depicts a sample diagnosed as SqCC. Figure 6, E and F, shows H&E and p40 images, respectively, whereas Figure 6, G and H, shows the annotation (T0) and whole-spot prediction (T2), respectively. Here, the regions predicted by CIRECA_SHP to be SqCC colocalize extremely well with the p40-positive and the annotated regions. This observation is different from the situation of ADC discussed above, where SHP predicted cancer in larger areas than those indicated by annotation and IHC. This might be an indication about the tendency of cancer infiltration into surrounding tissue.

Whereas the results in the previous section discussed examples of the vast majority of cases in which COH core histology, UMP, IHC, and CIRECA_SHP agreed, in this section, the discrepancies among these techniques will be analyzed. The percentage of COH-UMP discrepancies (15.5%) was, by number, about the same as the UMP-IHC discrepancies (16.3%; see below) and much higher than the disagreements between SHP results and UMP (see Tables 5 and 6). The accuracy of IHC was evaluated against UMP as a gold standard and is listed in Table 7. The good agreement between UMP annotation and SHP is, of course, partially because the SHP algorithm was trained on regions annotated by UMP. Nevertheless, it is highly encouraging to realize that the vast majority of SHP results are in agreement with IHC, even if the histology and pathology diagnoses disagreed.

The following discussion of these discrepancies is organized as follows. First, 14 cases diagnosed as AdSqCC by COH histology, but as either ADC or SqCC by UMP, will be discussed. It should be noticed that AdSqCC cases are listed as a separate class in Table 1. This discussion will also include cases for which CIRECA_SHP predicted AdSqCC but that were classified as either ADC or SqCC by pathology. Next, cases will be discussed where core histology suggested ADC and pathology suggested SqCC, or vice versa. In these cases, CIRECA_SHP generally agreed with the IHC results. Finally, cases will be discussed where core histology suggested normal tissue, but pathology suggested cancer, or vice versa. As pointed out before, the differences between UMP and COH diagnoses may be due to the fact that over the dimensions of the tumor or the tissue core, the actual cancer type changed. Thus, it is not a matter of one or the other diagnosis being incorrect, but rather that CIRECA_SHP presents a nondestructive method to analyze an actual sample on a TMA that was sectioned from a tissue core that may or may not have been representative of the tumor, or may have changed significantly over the dimension (length) of the core.

The diagnosis of AdSqCC appears to be a nontrivial issue given the fact that there was significant disagreement between COH histology and UMP: 14 cases with a COH core histology of AdSqCC were diagnosed by UMP as either ADC (6 cases) or SqCC (8 cases). In the literature, there are 2 histologic descriptions of AdSqCC: in one of them, referred to as biphasic, cells from areas that are clearly ADC and cells that are clearly SqCC merge at the margins of separate tumors and create regions of mixed cancer. The other description uses the term mixed (or admixed) AdSqCC, in which anaplastic tumor cells that may arise from multipotent stem cells show no microscopic evidence of squamous or glandular differentiation.³  Although the admixed designation is no longer recognized within the World Health Organization classification and such cases are designated as non–small cell lung cancer, not otherwise specified, it is interesting to point out that this SHP-based study found and distinguished 2 different types of AdSqCC because of its sensitivity to detect heterogeneity, as discussed below.

Table 8 lists the AdSqCC samples in the data set examined here, along with the UMP annotation and the CIRECA_SHP results. The gold standard column is based on a combination of UMP and IHC response. CIRECA_SHP agrees with the gold standard except in 2 cases, CHLC02_1109 and CHLC04_0509, indicated with a footnote designator in the rightmost column of Table 8. These 2 samples had positive p40 response and are considered false SHP results.

Four representative case studies from these 14 discrepancies are depicted in Figure 7, A through L, where 3 cases are presented in which CIRECA_SHP and IHC suggest either ADC or SqCC instead of the AdSqCC histology diagnosis. In contrast, Figure 8, A through D, presents 1 of 2 cases in which CIRECA_SHP and IHC suggest mixed AdSqCC, although histology and pathology favor either ADC or SqCC.

In Figure 7, the left column (Figure 7, A, D, G, and J) shows either H&E-stained tissue sections (Figure 7, A) or positive IHC results (Figure 7, D, G, and J). The second column (Figure 7, B, E, H, and K) depicts SHP prediction (T2) images displayed in the same color scheme as discussed before. The third column (Figure 7, C, F, I, and L) depicts the T0 annotation images. It should be noted again that the annotation images of a given spot were not included in the training of the prediction algorithm. Because this study was carried out in a LOOCV mode, all samples were analyzed in a blinded fashion by the classifier trained on all other data. However, for the final evaluation of the accuracy of the method, the true or gold standard result needed to be available; hence, annotation images were obtained and are presented here.

Figure 7, A through C, shows results for a tissue spot with a core histology of AdSqCC that was diagnosed by UMP as ADC. Immunohistochemistry was inconclusive in this case because neither TTF-1 nor p40 showed positive response. CIRECA_SHP classified the tissue as ADC (Figure 7, B) and cancer-adjacent normal tissue, in agreement with the UMP annotation (Figure 7, C). The regions annotated as necrotic tissue (12-o'clock position) in Figure 7, C, were not identified as necrosis with high confidence, and consequently are shown in white in Figure 7, B.

Similarly, Figure 7, D through F, depicts a sample with core histology of AdSqCC that was annotated by UMP as SqCC (Figure 7, F). This sample showed weakly positive p40 IHC response (Figure 7, D) and was classified as SqCC by CIRECA_SHP. In this case, and in many of the other discrepancies, the CIRECA_SHP results agreed with the IHC response. However, it should be emphasized that IHC results were not available when the original COH core histology diagnosis was rendered.

The next 2 rows in Figure 7 represent 2 core samples from the same patient. This case study demonstrates the power and clinical utility of CIRECA_SHP because it can detect, in a completely nondestructive way and without the use of an external label, the exact pathology of a sample that may be in a clinical trial or otherwise of high importance. The third- and fourth-row samples (Figure 7, G through I and J through L) are from the same patient and presumably from the same tumor resection. Both cases were diagnosed as AdSqCC by COH core histology, but responded very differently to IHC. Figure 7, G, shows strong p40 response and was independently diagnosed as SqCC and adjacent normal tissue (Figure 7, I) by UMP. CIRECA_SHP concurred with the UMP diagnosis, and the regions identified as SqCC by UMP colocalize very well with the p40 response and the SqCC classification by CIRECA_SHP. The results depicted in Figure 7, J through L, are from the same patient. Figure 7, J, exhibited strong TTF-1 response, and was diagnosed by UMP and CIRECA_SHP as ADC with adjacent normal tissue areas (Figure 7, L). Thus, samples CHLC13_0804 and CHLC13_0805, which are from the same patient, represent an example of biphasic AdSqCC where the 2 cores represent contiguous regions of one or the other subtype. In this particular set of cores, different outcomes were found by SHP and confirmed by IHC that may confound any clinical trial because of the inherent variability of tissue samples.

In the entire data set of 328 samples, there were only 2 samples that exhibited mixed ADC/SqCC spectral signatures and were classified as AdSqCC by CIRECA_SHP. An example of these is shown in Figure 8. This sample was diagnosed as ADC by COH core histology, but as SqCC by UMP (Figure 8, B). This sample showed weak TTF-1 and weak p40 response (Figure, 8, A and C) and a near 50:50 mixture of ADC and SqCC by CIRECA_SHP (Figure 8, D). The weak TTF-1 and p40 responses, the contradictory histology (ADC) and pathology (SqCC), and the admixed ADC and SqCC SHP response suggest that this sample is indeed AdSqCC. This example illustrates that some tumor cores may exhibit enormous tissue heterogeneity that CIRECA_SHP can resolve. It is interesting to note that the areas predicted to be SqCC by CIRECA_SHP agree quite well with the SqCC annotation regions, and that ADC was predicted mostly in the unannotated regions.

Whereas the patient therapeutic options are not affected by a false-positive or false-negative diagnosis of AdSqCC, misdiagnosis of ADC versus SqCC, or vice versa, has major health implications for the patient. This is because modern therapies are targeted and designed to block specific cancer oncogenes, or, in the case of the drug bevacizumab, an endothelial growth factor receptor. Bevacizumab is used for the treatment of ADC but is contraindicated in SqCC because of possibly fatal side effects.⁴² 

In the data presented here, there were 18 cases with discrepant cancer type histology and pathology. These cases are summarized in Table 9. There were 10 patients with COH histology of SqCC that were diagnosed as ADC by UMP. All these patients had positive TTF-1 response and therefore one may assume that at the level of the tissue section analyzed, the UMP diagnoses were correct. Eight of these 10 patients were classified as ADC by SHP, with the other 2 giving mixed ADC/SqCC results. Similarly, there were 7 patients with COH histology of ADC and UMP diagnoses of SqCC. Five of these showed positive results for p40 IHC, and all 5 were classified as SqCC by SHP. Examples of these discrepancies are shown in Figure 9. Finally, 1 patient was classified by COH pathology as non–small cell lung cancer and ADC by UMP. This sample was negative for both p40 or TTF-1 immunostains but was clearly classified as ADC by SHP.

In Figure 9, the top 2 rows (Figure 9, A through C and D through F) depict cases that had COH histology diagnoses of SqCC, but UMP annotation of ADC with cancer-adjacent normal tissue (Figure 9, C and F). Both these samples showed positive TTF-1 response (Figure 9, A and D) and clear ADC classifications by CIRECA_SHP (Figure 9, B and E). Similarly, the 2 bottom rows (Figure 9, G through I and J through L) had COH histology diagnoses of ADC but UMP annotation of SqCC with cancer-adjacent normal tissue (Figure 9, I and L). The p40 immunostain was positive for tumor cells in these 2 samples (Figure 9, G and J) and CIRECA_SHP clearly classified these samples as SqCC (Figure 9, H and K).

These results reemphasize the well-known fact that purely morphology-based ADC-SqCC discrimination is less reliable than pathology augmented by IHC, but further demonstrate that the inherent compositional (biochemical) signatures detected and used by CIRECA_SHP for classification provide enhanced reliability and reproducibility over classical methods of histopathology. The changes in signatures due to compositional differences are observed not only among cancerous tissues, but in normal tissue types as well. In fact, large spectral changes due to the presence of compounds such as glycogen (in cervical tissue⁴³  and liver⁴⁴ ) or collagen in bone⁴⁵  or scar tissue¹  have been reported (see also Figure 1). Furthermore, normal glandular and squamous tissue are easily distinguished by their different spectral signatures due to different biochemical composition. This was demonstrated in early SHP studies of endocervical and ectocervical biopsies.²⁴ 

There was a total of 12 cases where the core histology (COH diagnosis) indicated either ADC or SqCC, but pathology and SHP indicated proximal normal tissue. Here, it is possible that the tissue sections from which the core histology was rendered at COH were sufficiently different from sections A001 through A003 (see above) used in this study, and cancerous cells may not have been present in the sections investigated in our study. Nevertheless, the agreement between UMP and SHP was encouraging: of 9 patients with ADC histology and normal pathology, 7 showed only proximal normal and macrophage response by SHP. CIRECA_SHP indicated ADC for 1 sample, but with low confidence. Finally, SHP gave a false-positive result for 1 sample. An example from these 9 samples is shown in Figure 10, A through C. This sample had a COH histology diagnosis of ADC but a UMP annotation of normal/MAC, as shown in the annotation (T0) image, Figure 10, C. The CIRECA_SHP whole-spot prediction (T2) image, Figure 10, B, showed mostly proximal normal and smaller areas of distal normal (red). This mixture of proximal and distal normal tissue indicates the likelihood of this tissue section being relatively close to a cancerous lesion. Thus, it is likely that the original tissue block from which the core was obtained had, in fact, cancerous tissue that was not represented on this section.

Additionally, there were 3 patients with SqCC diagnosis by COH and normal diagnosis by UMP. In one of them, SHP did not detect any cancer, in agreement with pathology. In the patient sample shown in Figure 10, D through F, positive TTF-1 response was observed in (possibly) benign alveolar cells, shown in Figure 10, D. However, the regions of positive TTF-1 results were annotated at UMP as normal tissue (Figure 10, F), whereas SHP detected some abnormalities in these regions, predominantly ADC. It is not clear, at this point, whether or not the regions with positive TTF-1 response were actually benign, and CIRECA_SHP misclassified the TTF-1 positive regions as cancer, or CIRECA_SHP detected the slight biochemical changes associated with incipient ADC but morphologic changes were not sufficient for the pathologist to classify it as ADC.

Finally, Figure 10, G through I, demonstrates that even very small cancerous regions detected by pathology (Figure 10, I, arrow) are reproduced in the SHP prediction image, Figure 10, H. The detection via CIRECA_SHP of small foci of cancer, or micrometastases in tissues such as lymph nodes,^46,47  will open new avenues in rapid and highly accurate screening for cancerous lesions.

The results presented herein demonstrate that CIRECA_SHP delivers a level of diagnostic accuracy that at least matches that of classical histopathology combined with IHC. This very high accuracy results from the use of inherent, optical signatures in SHP, which are manifestations of the biochemical composition of tissue pixels. These signatures can be observed with a very high degree of reproducibility.

The use of multivariate mathematical analysis transforms the observed raw spectral data sets into images that depict heterogeneity in a tumor, tumor types and subtypes, the effect of a cancerous lesion on its surroundings, and the presence of tumor-infiltrating immune cells. Furthermore, CIRECA_SHP detects slight differences in chemical composition in samples from different patients with the same type of lung cancer (ADC or SqCC). Future studies are needed to determine whether these findings can help identify cohorts of patients with similar outcome to the same therapy and use this information prognostically.