Spectral discrimination of breast pathologies in situusing spatial frequency domain imaging

A compact SFDI system (purchased from Modulated Imaging Inc., Irvine, CA) illuminated breast surgical-resection specimens with a harmonically modulated, planar source at four NIR wavelengths (658, 730, 850, and 970 nm) [37]. Structured light patterns were projected onto each tissue surface at 30 spatial frequencies uniformly distributed between 0 mm^-1 and 0.33 mm^-1 by using high-power light-emitting diodes (LEDs), a projection system, and a digital micro-mirror device. The projector and camera subsystems were described in previous publications [37, 39] and were fully integrated in a portable platform mounted on a z-axis post. In total, 360 images were acquired per tissue (30 spatial frequencies × 3 phase offsets × 4 wavelengths) in approximately 10 minutes. Data were simultaneously acquired over the full field in a noncontact geometry, in which the acquisition field of view (FOV) was determined by magnification of the illumination and collection optics, here optimized to image a 5.5-inch × 7.5-inch area. A 12-bit CCD-based camera, co-registered with the projector, collected diffusely reflected light at a 696 × 520 pixel resolution with 2 × 2 pixel binning (full pixel resolution of the camera is 1,392 × 1,040). Fundamentally, however, the detected spatial resolution is limited by the physics of light transport in tissue, and contrast-detail analysis was previously performed to evaluate the minimized size of detectable scattering contrast in tissue-simulating phantoms [40]. A three-phase demodulation scheme was then used to extract the frequency-dependent modulation amplitude and to remove noise and ambient light from the imaged field [41]. The spatial frequency of the illumination pattern integrates optical parameterization with effective sampling depth, in which higher spatial frequencies attenuate more rapidly and thereby sample superficial tissue volumes. The effective sampling depth is also determined by the tissue optical properties. Theoretically, probing depths of 1 to 8 mm are expected in tissues; experimentally, we demonstrated detection of scattering inclusions up to depths of 1.5 mm in tissue-simulating phantoms by using the higher spatial frequencies used in this study [40]. At each imaging session, the modulation amplitude of a Siloxane titanium dioxide (TiO₂) reflectance standard with known optical properties was imaged to calibrate spatial nonuniformity in the illumination and imaging systems. A 1-mm uniform height offset was erroneously, but systematically, introduced between measures of the tissue and reference standard. The offset did not affect tissue discrimination and resulted in negligible scaling of the modulation amplitude because reflectance decays with distance, h, according to an inverse square law (the distance between the CCD lens and tissue was ~120 mm h 2 2 / h 1 2 = 119 m m 2 / 120 m m 2 ≅ 1[42].

Optical parameters were estimated by minimizing the residual sum of squares between the measured modulation amplitude and its forward simulation, according to a scaled Monte Carlo model of light transport [37]. Monte Carlo solutions were stored in a look-up table (LUT) to reduce the computational burden associated with iterative estimation of spectral parameters [39]. Subsequently, spectral parameters were fit according to absorption by the endogenous tissue chromophores oxygenated-hemoglobin, deoxygenated-hemoglobin, and water, and the wavelength-dependence of light scattering, and produced estimates of scattering amplitude (A), scattering slope (b), total hemoglobin concentration (HbT), percentage hemoglobin oxygenation (%O ₂ ), and percentage water (%H ₂ O). Data-acquisition and analysis methods are discussed more comprehensively in a companion article [39].

The ability of SFDI to distinguish histopathology, here used as the diagnostic gold standard, was evaluated at the cut surfaces of 47 surgical breast tissues. In this HIPAA-compliant, prospective study, approved by the Institutional Review Board for the protection of human subjects at Dartmouth, written informed consent was not required for participants, although an Information Sheet detailing the study was provided with an opt-out provision. All of the study patients were female. A breakdown of the pathologic characteristics of the imaged cancers is provided in Table 1. In addition to these 27 biopsy-proven cancers, 11 fibroadenomas and nine normal (including fibrocystic disease) breast cases were imaged. At Dartmouth, the breast-resection specimens are inked with six pre-assigned colors to preserve orientation within the breast, so that, if it is found by pathologic evaluation that the margins are involved by tumor, the patient is called back to undergo a second resection of those specific margins rather than the entire surgical cavity.

The fresh, inked specimens, from patients who did not decline this research use of their tissues, were transported immediately to the Department of Pathology. By standard protocol, all of the specimens were “bread loafed” in their entirety into consecutive slices up to 0.5 cm in thickness, and, for this study, one face of one slice representing the lesion and its relation to the nearest margin was imaged. No surgical margin was imaged en face, given the required orientation inking by the surgeon. The exact area of tissue imaged in one slice was outlined with four inked pins pushed through the tissue block. The tissue was returned to Pathology after imaging, within 30 minutes of resection, for standard histologic processing. The imaged slices were formalin-fixed with the pins in situ before being removed for usual processing. The tiny inked pinholes survive processing and are evident in the histologic sections of that imaged tissue, facilitating optimal registration with the histologic findings. The effects of tissue shrinkage after processing and sectioning distortion were minimized by ensuring optimal formalin concentration and fixation time in accordance with College of American Pathologist (CAP) best practices. Specimen imaging did not affect procedure time in the operating room, processing time in the Department of Pathology, or the content and time to verification of the final pathology report.

The findings in histologic sections representing the imaged areas were included in the final pathology report. Figure 1 illustrates co-registration between photographs of the cut face of one slice in the specimen, with the lesions outlined in diagnosis-defined colors (row 1), corresponding histopathology (row 2), and spectral images (rows 3 to 7). An experienced breast pathologist (WAW) outlined tissue areas containing target breast lesions on digital photographs of each specimen, guided by co-registered hematoxylin and eosin (H&E)-stained tissue sections from the imaged surface. Affine transformations were manually performed in Photoshop to convert the outlined tissue areas into masks for region-based image analysis in MATLAB (The MathWorks Inc., Natick, MA, USA).

In total, 48 specimens from 47 patients were consecutively imaged in this study; one was excluded from analysis because surgical inks bled into the primary tissue field. Surgical inks, sometimes observed at the tissue edge, were avoided during region of interest (ROI) selection and masked according to the R² coefficient of determination (pixels excluded if R ² < 0.95). Spatial frequency- and wavelength-dependent reflectance measurements were evaluated on a pixel-by-pixel basis in 59 regions of interest (>265,000 spectral pixels). Discrimination was assessed between benign and malignant pathologies and between the benign pathology subtypes (normal and fibroadenoma), and the malignant pathology subtypes (DCIS, invasive cancer, and partially treated invasive cancer after neoadjuvant chemotherapy), as listed in Table 2. In this analysis, fibrocystic disease was grouped with normal breast pathologies. Tumors treated with neoadjuvant chemotherapy before surgical resection (n = 5) were imaged to observe how chemotherapy alters tissue optical properties, the better to inform diffuse tomographic monitoring of treatment response to therapy [43].

Spectral maps were associated with morphologic and immunohistochemical markers identified in adjacent, stained sections of the tissue, cut in the exact geometry imaged in situ. Possible correlates between immunohistochemical measures and spectral parameters were evaluated according to the Pearson correlation coefficient. CD31 (platelet endothelial cell adhesion molecule) was used to stain immunohistochemically the preexisting endothelial cells as a general indicator of normal vasculature, even though malignant vessels have been reported to retain this antigen [44]. CD105 (Endoglin), thought to antagonize the inhibitory effects of transforming growth factor-beta (TGF-β) on cell proliferation and capillary formation [44], was used to stain immunohistochemically a transmembrane glycoprotein expressed predominantly on the surface of highly proliferating endothelial cells, as an indicator of tumor angiogenesis. Cytokeratin CK5D3 was used as a marker for breast epithelium (benign and malignant). All tissue was fixed in 10% buffered formalin (Biochemical Science Inc, Swedesboro, NJ, USA), dehydrated through graded alcohols, and paraffin embedded. Tissue sections (4 μm) were coated with adhesive (Sta-on; Surgipath Medical Industries, Richmond, IL, USA), mounted on glass slides, and stained with H&E for initial review. Tissue sections were air-dried for at least 30 minutes and then loaded onto a Leica Bond Max automated immunostainer. Here, the sections were baked (30 minutes at 60°C), dewaxed for 30 minutes at 72°C, rinsed with alcohol, and washed in Bond wash buffer. Antigen retrieval was accomplished with Bond epitope Retrieval 2, Ar9640 (pH 8.9 to 9.1) for 20 minutes at 100°C; followed by cooling for 12 minutes and a rinse in Bond wash buffer. All dewaxed, antigen-retrieval, and detection reagents were supplied by Leica Microsystems (Bannockburn, IL, USA). The primary antibodies were incubated for 15 minutes: CK5D3 at 1:100 (Biogenix, Fremont, CA, USA), CD31 at 1:50 (Dako, Carpinteria, CA, USA), and CD105 at 1:60 (Vector, Burlingame, CA, USA). Diaminobenzidine was applied for visualization with a hematoxylin counterstain.

Using Media Cybernetics™ Image Pro (Media Cybernetics, Bethesda, MD, USA) and automated stage control bundled software (Surveyor© Automated Specimen Scanning stage control bundled software (Objective Imaging Ltd., Cambridge, UK), whole H&E-stained slides and CD105 and CD31-immunostained slides, representing the imaged areas of lesional breast tissue, were digitally scanned at high resolution and montaged. Mean vessel density (MVD) was quantified according to the areas of CD31-positive (preexisting vasculature) or CD105-positive (tumor-induced) vessels, and segmented in pseudo-color, as a percentage of the total slide area. Mean vessel area (MVA) was determined from the combined areas of CD31-positive or CD105-positive blood vessels, segmented in pseudo-color, and measured in square micrometers. The epithelial component of each imaged area was defined as the percentage of the total slide area showing positive immunostaining for CK5D3 and segmented in pseudo-color. On the same slide, the morphologically distinctive white spaces representing lipocytes (emptied of lipid as a result of fixation and tissue processing) were also segmented in pseudo-color as a percentage of the total slide. The remaining percentage of the slide represented connective tissue stroma. The diagnostic ROIs were selected by WAW, a board-certified surgical pathologist with 15 years of expertise in breast pathology and image analysis. The image thresholding and processing were performed by MCS, a technologist with 6 years of expertise in the Image Pro software.

A nearest-neighbor learning algorithm was implemented to automate spectral discrimination of benign and malignant pathologies, and all pathology subtypes listed in Table 1; performance was evaluated by using a threefold cross-validation [45]. All data were randomly divided into three nonoverlapping sets, with an equal number of reflectance spectra per diagnostic class per set. Two of these sets were used for training, and the third set was used for validation. Training pixels were associated with a known diagnosis, according to the pathologist’s demarcation of lesions; the diagnosis of each validation point was also identified by the pathologist, but remained blind to the classifier to evaluate its performance. Classification error was taken to be the percentage of misclassified pixels in the validation set, where a misclassification means that the diagnosis assigned by the classifier did not match the diagnosis provided by the pathologist. Performance metrics were quantified 3 times for all possible permutations of the training and validation sets, and reported values were the average of these three executions. The nearest-neighbor algorithm assigned each unclassified parameter set to the majority diagnosis of its k nearest neighbors found in the training space by an efficient k-dimensional tree-search algorithm [46]. A whitening transformation was applied to all spectral parameters before diagnostic discrimination to prevent amplitude weighting. Additionally, outliers were removed from the training set according to their interquartile fractions (comprising 5% of the total data set). Query points were never marked as outliers, because this information was not known a priori.

Receiver operating characteristic (ROC) analysis was used to optimize classifier sensitivity and bias as a function of nearest-neighbor number, k. Confidence intervals (α = 0.05) for the sensitivity and specificity were computed according to the Yates χ² interval [47].

An iterative search algorithm, sequential floating forward selection (SFFS), was implemented to rank the contribution of each spectral parameter to tissue-type discrimination. The Bhattacharyya statistical distance, J _ij , was used to measure the separation between two diagnostic classes (i,j) [48, 49] and was generalized to all spectral parameters (n = 5) by:

Here, P_i represents the prior probability of class i determined by its fraction of pixels in the training set. The search algorithm ranked all parameters in order of the contribution to diagnostic-class separability.

Examples of optical maps co-registered with digital photographs of the surgically resected tissues are shown in Figure 1 for each lesion type evaluated. Boxplots in Figure 2 quantify the extensive distribution of optical and immunohistochemistry markers for all reflectance measures evaluated on a pixel-by-pixel basis (>265,000 measurements); mean values per patient are plotted as green dots that overlie the boxplots. Median values were compared to identify optical signatures specific to each diagnostic class, and interquartile fractions reflect the inherent heterogeneity observed within and between tissues. Median values were significantly different within 95% confidence limits for nonoverlapping notches. Distributions of the scattering amplitude and scattering slope showed less variance than functional absorption parameters, like hemoglobin concentration and oxygenation status, because the morphologic transformations that determine light scattering occur on submicron and even subnanometer length scales [50, 51]. Absorption parameters typically signal broader functions, and their distributions were characterized by greater variance because of surgery-induced artifacts like vascular disruption, eradication of the tissue’s oxygen supply, and eventual tissue dehydration. Efforts were made to image all tissues within the same time window and to avoid sampling of cauterization and hemorrhage, but limited artifacts are plausible. Surgical inks at the edge of the resected specimens may have also contributed to outliers with artificially high hemoglobin concentrations. This variability was also detected immunohistochemically in the subsequently fixed tissues, as shown in Figure 2f. However, SFDI assess the full tissue surface, accounting for inherent biologic heterogeneity, so that on average, functional biomarkers may strengthen tissue-type discrimination.

The distributions in Figure 2 show that observed hallmarks of normal or fibrocystic pathologies included a heightened scattering amplitude and slope, as compared with malignant pathologies. Lower scattering amplitude was observed in fibroadenomas, which contain clearly encapsulated and abundant stromal cellularity, confirming diagnostic trends established in previous studies of localized spectroscopic scattering from breast tissues [16]. Higher hemoglobin and reduced oxygenation values were observed in invasive cancers, as compared with the benign pathologies. DCIS was spectrally distinguished from other pathologies by decreased oxygenation and increased water levels, suggesting heightened metabolism and a proportional propensity for more surrounding fibrous stroma. However, a small DCIS patient population (n = 4) and spectral parameterization limited interpretation of this signature (fits were based on the extinction of pure, unbound water and did not include absorption by adipose tissue). Neoadjuvant chemotherapy increased the scattering slope when comparing its untreated, invasive counterparts, regardless of response to therapy. An increase in scattering slope suggests cell shrinkage, given its increased propensity for Rayleigh-type scattering.

Increased oxygenation levels were also induced in tumors partially responsive to chemotherapy and could be valuable indicators for therapy monitoring. Many aggressive tumors thrive in hypoxic environments [52], so increased oxygenation levels suggest a return to the normal phenotype.

Pairwise correlations between spectral parameters and quantitative immunohistochemistry are summarized in Table 3. Collagen in the extracellular stroma demonstrated a strong, positive correlation with scattering slope because the small fibrils that form collagen fibers further enhance Rayleigh-type scattering [53]. In contrast, a strong negative correlation was observed between tissue epithelial content and the scattering slope. Malignant tissues expressed greater levels of angiogenic vessels and preexisting vascular endothelium. This net growth in vascularity was detected spectroscopically by increased total hemoglobin, but no optical distinction was made between CD105-specific and CD31-specific vasculature. Immunohistochemical measures were sampled in discrete locations about the tissue surface, limiting its complete representation of ROIs.

Spatial frequency- and wavelength-dependent reflectance measures were applied to the discriminant classifier on a pixel-by-pixel basis, and separation of all tissue subtypes was achieved with 82% accuracy. Figure 3a shows that diagnostic sensitivity increased with nearest neighbor number at the cost of specificity or bias. ROC curves revealed that 11 and nine nearest neighbors optimized separation of benign and malignant pathologies and all pathology subtypes listed in Table 1, respectively. The confusion matrix in Figure 3b explicitly reflects the true and predicted diagnosis as a percentage of total number of diagnosed pixels. True positives per tissue type are along the diagonal, and misclassifications are off-axis; showing that, for example, 16% of treated invasive tumors were misclassified as normal. Figure 3c through e shows a false-color map of the multiparametric diagnosis co-registered with pathology for a malignant and benign tissue to illustrate application of the classifier to a patient diagnosis. Diagnostic performance is summarized in Table 4, which reports the sensitivity, specificity, positive predictive value (PPV), negative predictive value (NPV), and accuracy observed per tissue type. Discrimination of benign from malignant pathologies was highly specific (93%) and reasonably sensitive (79%), although sensitivity was limited in underrepresented pathologies like DCIS and residual invasive cancers after neoadjuvant chemotherapy (61% to 66%). Discrimination between benign and malignant pathologies was more accurate (88%) because these distribution sizes were more equally represented in the classifier training set. Ultimately, the surgeon can moderately regulate diagnostic sensitivity and specificity by varying the number of nearest neighbors used by the classifier to generate a diagnosis, as shown in Figure 3a. The NPV (89%) between benign and malignant pathologies was higher than the PPV (86%). In practical terms, surgeons are more interested in a high NPV when high confidence exists that no residual cancer has been left behind. A lower PPV, correlating with a false-positive call on any margin, would mean that the surgeon would take a little more tissue at the appropriate margin at the time of the first surgery. Even if that margin does not subsequently turn out to be positive, this can be justified clinically by the improved cosmetic outcome, the accuracy of immediate margin reexcision, the reduced costs of a single surgical procedure, and patient well-being by avoiding a second, separate reexcision procedure. High NPVs were also reported for underrepresented tissue types, but these values appear inflated when the number of true negatives significantly outweighed the number of true positives.

The SFFS algorithm ranked the region-averaged (1) scattering amplitude, (2) percentage water, (3) total hemoglobin concentration, (4) the scattering slope, and (5) percentage oxygen, as most significant to tissue discrimination in that order. The scattering amplitude was most valuable to pathology discrimination, even through qualitatively, the scattering slope appeared to localize better with suggestive fibroglandular lesions. The scattering slope was potentially undervalued by the limitations of pathology co-registration. Percentage oxygenated hemoglobin was least valuable to a diagnosis, mainly because of ex vivo fluctuations in hemoglobin oxygenation status. Bydlon [54] showed that percentage of oxygenated hemoglobin varies nonlinearly within 30 minutes of resection, but other spectral parameters, like the scattering coefficients and total hemoglobin concentration, remained relatively stable after excision.