2.1.

Samples and Histological Data

Normal and tumor breast samples were collected from 80 patients with invasive ductal carcinoma following lumpectomy from St. Joseph’s Hospital, Hamilton, Ontario, Canada [Hamilton Integrated Research Ethics Board (Project No. 10-3393)]. Tumor samples were collected near the center of the tumor; whereas, normal samples were collected from the exterior of the excised tissue. Samples were frozen and stored at $-{80}^{°}\mathrm{C}$. A study by Shalaby et al.) showed that freezing and thawing normal and tumor breast tissue did not result in significant differences in optical properties.²⁷ Prior to measurement, samples were thawed and trimmed to fill a sample holder with a 12.7 mm diameter and 1 mm deep cylindrical sample holder. The holders are covered on both sides using optically transparent ultralene film to hold the samples in place and flatten the tissue surface.

After measurement, samples were prepared for histological analysis. Samples were fixed in formalin, alcohol washed, and embedded in paraffin to form tissue blocks. Tissue blocks were sectioned and mounted onto glass slides for H&E staining. Following staining, the glass slides were digitized and registered with optical raster scans to quantify 1 mm × 1 mm pixel divisions. Each pixel was evaluated by a pathologist to classify the tissue composition as percent tumor, fibroglandular, and/or adipose.

2.2.

Instrumentation

Mounted samples were placed in the BEAR system for measurement.⁴⁴ This combined system uses one 3 mm diameter custom probe, as shown in Fig. 1, to make spot measurements. The probe was positioned $\sim 1\mathrm{mm}$ from the surface of the tissue for each sample measured. At the center of the probe is a $400\mu \mathrm{m}$ diameter fiber used for TRF excitation using a 355 nm Nd:YAG laser (PNV-M02510-130, Teem Photonics, France) and detection. An acousto-optical tunable filter (TEAF5-0.36-0.52-S-MSD, Brimrose, United States) is used to measure fluorescence wavelengths from 380 to 570 nm in increments of 5 nm. The laser is triggered to send 11 pulses per wavelength with a power of $\sim 3\mathrm{mW}$ at the tissue surface. Fluorescence is measured using a photomultiplier tube (C5594-12, Hamamatsu Photonics, Hamamatsu City, Japan) biased to 2000 V. The output was then sampled using a digitizer (ADQ7DC, 14 bit, 10 GS/s, 2.5 GHz bandwidth, SP Devices, Linköping, Sweden). The laser is pulsed to make 11 measurements per spot.

Fig. 1

Geometry of the 3 mm diameter measurement probe containing a $400\mu \mathrm{m}$ diameter TRF excitation and collection fiber, a $200\mu \mathrm{m}$ DR excitation fiber, and three detection fiber bundles, 0.32 mm, 0.64 mm, and 1.5 mm from the DR excitation fiber.

The probe also contains a $200\mu \mathrm{m}$ diameter fiber used for DR excitation using a broadband Xenon lamp (ASB-XE-175BF, Spectral Products, Putnam, Connecticut, United States) as well as multiple $200\mu \mathrm{m}$ diameter fibers used for reflectance detection. These DR detection fibers are bundled such that there are three detection bundles 0.32 mm, 0.64 mm, and 1.5 mm from the DR excitation source fiber. Each bundle is connected to a separate spectrometer (BW-UVIS, 600 g/mm, $50\mu \mathrm{m}$ slit, StellarNet, United States) to record reflectance.

Samples were placed on an $x\text{-}y$ linear stage (XLS-1-80-1250, Xeryon, Leuven, Belgium) to translate the samples for raster scanning. Samples were translated in $1\mathrm{mm}\times 1\mathrm{mm}$ step sizes and a spot measurement was performed at each location. Figure 2 shows a photograph of a sample and its corresponding histology slide with grids overlaid depicting the raster scan. An averaged spectrum for one of the spot measurements and the corresponding region in the histology image is shown below. Raster scanning enabled rapid measurement of the entire sample surface, allowing for a greater number of spectra to be collected in a shorter time period. All components were connected to a computer to automatically control the timing and execution of sample excitation, measurement, stage translation, and recording.

Fig. 2

Photograph of the sample (a) and histology slide (b) of sample 11T. Grid overlays depict raster scan locations of each spot measurement. The averaged signal for some wavelengths of interest (400 nm, 460 nm, and 540 nm) for spot $x=4$, $y=8$ (c) and its corresponding histology image used for tissue classification (d) are also shown.

Each spot measurement takes approximately 1 s. The measurement time increases during raster scanning as spectra from each spot need to be saved and the translation stage requires time to move to the next position. In addition, the acquisition software performs coarse analysis to generate live-time data during the scan. As such, the time for each spot measurement increases to approximately 5 s during raster scanning. The speed of the imaging is mostly dependent on the step size of the raster scan. For a $13\mathrm{mm}\times 13\mathrm{mm}$ area raster scanned with $1\mathrm{mm}\times 1\mathrm{mm}$ step sizes, the measurement time is approximately 15 min.

2.3.

Data Processing and Analysis

Following measurement, the measured signal at each wavelength was averaged for the 11 laser pulses. Custom MATLAB (v 9.12; MathWorks, Natick, Massachusetts, United States) code was developed to fit the average spectra to extract the fluorescence intensity and lifetime of biomolecules of interest. The biomolecules of interest and their emission wavelengths are reported in Table 1. Fluorescence intensities of all emission wavelengths for each biomolecule were averaged and then normalized to the NADH intensity at 460 nm, the mean emission wavelength of NADH for each pixel. As such, the reported fluorescence intensities are collagen/NADH and FAD/NADH ratios. This normalization accounts for slight variations in sample to detector distance between measurements. In total, five TRF parameters were quantified: collagen/NADH intensity, FAD/NADH intensity, collagen lifetime, NADH lifetime, and FAD lifetime.

Table 1

Biomolecules of interest and emission wavelengths with 355 nm excitation.

A sample filled with de-ionized water was measured as background and subtracted from reflectance measurements collected from the second DR detection fiber bundle (0.64 mm). The reflectance data were then normalized to a 99% DR standard (Labsphere Inc., North Sutton, New Hampshire, United States). Optical phantoms made of India Ink (Higgins Ink, Leeds, Massachusetts United States) and $1\mu \mathrm{m}$-diameter polystyrene microspheres (Polysciences, Warrington, Pennsylvania, United States) were measured, and the reflectance data were used to create a look-up table for the extraction of the absorption coefficient, ${\mu }_a$, and reduced scattering coefficient, ${\mu }_s^{\prime }$, following the methods outlined by Le et al.⁴⁶ This look-up table allowed for the ${\mu }_a$ and ${\mu }_s^{\prime }$ at each wavelength to be extracted using the measured reflectance of each measurement. In total, eight DR parameters were quantified: normalized reflectance at 520 nm, normalized reflectance at 560 nm, ${\mu }_a$ at 540 nm, ${\mu }_a$ at 560 nm, ${\mu }_a$ at 576 nm, ${\mu }_s^{\prime }$ at 540 nm, ${\mu }_s^{\prime }$ at 560 nm, and ${\mu }_s^{\prime }$ at 576 nm. Greater detail regarding the selection of these parameters is outlined in Dao et al.⁴⁴

Multivariate analysis was performed using Orange software (v 3.32)⁴⁸ using extracted optical parameters from pixels classified as 100% tumor, fibroglandular, or adipose. In total, 761 adipose, 77 fibroglandular, and 347 tumor spectra were analyzed from 80 patients. Principal component analysis (PCA) was performed on three datasets: TRF parameters only, DR parameters only, and combined TRF and DR parameters. The number of principal components (PCs) was increased until it covered at least 98% of the variance in each dataset. There were four PCs used for the TRF and DR datasets and eight PCs in the combined dataset. The PCA loadings are included in the Supplementary Material. Decision tree modeling was also performed using Orange software. Classification was non-binary. The decision tree model classified tissue as adipose, fibroglandular, or adipose. The classifier was set to include a minimum number of three instances in leaves, not to split subsets smaller than five, and to stop when the majority reaches 95%. Tissue classification was performed using “leave-one-patient-out” cross-validation where the data from all but one patient is modeled and the data from one patient is tested and scored using this model. This process is then repeated for all patients in this study. Following testing, the receiver operating characteristic (ROC) curve was generated using the predictions from all folds.

To determine whether modeling with combined optical parameters improved tumor classification compared to either technique alone, McNemar-Bowker tests were conducted. The classification of the tumor pixels were compared between modeling using just TRF parameters and modeling using the combined TRF-DR parameters. Similarly, a second test was conducted to compare classification of tumor pixels using just DR parameters and the combined TRF-DR parameters.

3.1.

Time-Resolved Fluorescence Only

The fluorescence intensities of collagen and FAD as well as the fluorescence lifetimes of collagen, NADH, and FAD were used to create the TRF dataset. Four components were used to cover 99.2% of the variance in the TRF dataset. The two PCs covering the greatest variance, PC1 covering 43% of the variance and PC2 covering 34% of the variance, are plotted in Fig. 3. While the PCA showed some separation between tumor and adipose tissue in PC1, there is significant overlap between fibroglandular and the other tissue types. PC2 did not show much separation between tissue types.

Fig. 3

Results of PCA on TRF dataset. PC1 and PC2 are plotted for adipose, fibroglandular, and tumor tissue.

Tissue classification was performed using decision tree modeling with only the TRF parameters. The results are shown in Table 2. The ROC curve for tumor classification is displayed in Fig. 4. The sensitivity, specificity, and area under the ROC curve (AUC) of classifying each tissue group are reported in Table 3.

Table 2

Tissue classification using decision tree modeling with TRF parameters only.

Fig. 4

ROC for tumor classification using TRF dataset.

Table 3

Sensitivity, specificity, and AUC of classification of adipose, fibroglandular, and tumor samples using TRF parameters.

Modeling using only TRF parameters proved successful for adipose with a sensitivity of 90.8%, specificity of 77.6%, and AUC of 0.882. Tumor classification had a high specificity of 88.3%, but lower sensitivity and AUC of 72.3% and 0.842, respectively. TRF parameters did not classify fibroglandular tissue well, with a very low sensitivity of 14.3%. Therefore, the classification ability of TRF alone would allow for tissue classification in general but would not provide greater insight on the normal tissue.

3.2.

Diffuse Reflectance

The normalized reflectance at 520 nm and 560 nm as well as the ${\mu }_a$ and ${\mu }_s^{\prime }$, at 540, 560, and 576 nm were used to create the DR dataset. Four components were used to cover 98.5% of the variance in the DR dataset. The two PCs covering the greatest variance, PC1 covering 59% of the variance and PC2 covering 32% of the variance, are plotted in Fig. 5. The PCA using DR parameters showed greater clustering of adipose and tumor samples; however, the fibroglandular tissue showed greater variability within the group. PC1 showed greater separation of tissue groups than PC2, with PC 2 showing very little difference between tumor and adipose.

Fig. 5

Results of PCA on DR dataset. PC1 and PC2 are plotted for adipose, fibroglandular, and tumor tissue.

Tissue classification was performed using decision tree modeling with only the DR parameters. The results are shown in Table 4. Figure 6 shows the ROC curve for tumor classification. The sensitivity, specificity, and AUC of classifying each tissue group are reported in Table 5. Modeling using DR parameters greatly improved the classification ability for all tissue types when compared to TRF modeling. The greatest improvement was the fibroglandular sensitivity increasing from 14.3% to 63.6%. Better classification of fibroglandular tissue resulted in significantly fewer incorrect classifications of tumor, improving its sensitivity as well. As such, DR parameters are shown to be more effective at tissue classification than TRF parameters.

Table 4

Tissue classification using decision tree modeling with DR parameters only.

Fig. 6

ROC for tumor classification using DR dataset.

Table 5

Sensitivity, specificity, and AUC of classification of adipose, fibroglandular, and tumor samples using DR parameters.

3.3.

Dual Time-Resolved Fluorescence-Diffuse Reflectance

The TRF and DR datasets were combined to form the dual TRF-DR dataset. Eight components were used to cover 98.9% of the variance in the combined dataset. The two PCs covering the greatest variance, PC1 covering 42% of the variance and PC2 covering 21% of the variance, are plotted in Fig. 7. The PCA using both TRF and DR parameters showed similar clustering of tissue groups, with most of the separation resulting from PC1.

Fig. 7

Results of PCA on combined TRF-DR dataset. PC-1 and PC-2 are plotted for adipose, fibroglandular, and tumor tissue.

Tissue classification was performed using decision tree modeling with both TRF and DR parameters. The results are shown in Table 6. The ROC curve for the combined dataset is shown in Fig. 8. The sensitivity, specificity, and AUC of classifying each tissue group are reported in Table 7. Modeling using both TRF and DR parameters further improved the classification ability for all tissue types.

Table 6

Tissue classification using decision tree modeling with combined TRF-DR parameters.

Fig. 8

ROC for tumor classification using the combined TRF-DR dataset.

Table 7

Sensitivity and specificity of classification of adipose, fibroglandular, and tumor samples using combined TRF and DR parameters.

Combining both optical techniques improves tissue classification compared to either technique alone, as shown in Table 8, summarizing the sensitivity, specificity, and AUC using each technique. Combining datasets did not improve the sensitivity of fibroglandular classification, but it did reduce the misclassification of some fibroglandular measurements as tumor from 20 false positives using the DR dataset only to 17 using the combined dataset.

Table 8

Summary of sensitivity, specificity, and AUC of positive margin (tumor) classification using each dataset.

For each tumor pixel in this study ($n=347$), the classification for each model was examined to determine whether combining TRF and DR parameters improved tumor classification. Table 9 shows tumor classification using only TRF parameters and how the classifications changed following modeling using the combined parameters compare. In 50 instances, tumor pixels that were incorrectly classified as adipose using modeling with only TRF parameters were then correctly classified as tumor using the combined dataset. Similarly, 23 of the tumor pixels incorrectly classified as fibroglandular using TRF parameters were then correctly classified using the combined parameters. However, for 27 tumor pixels correctly classified using only TRF parameters, 22 were mis-classified as adipose and 5 as fibroglandular using the combined dataset. Overall, combining the parameters resulted in 46 more correctly classified pixels, a 13.3% increase. With a McNemar–Bowker test $p$-value of $<0.01$, there is a statistically significant difference between the two classification models.

Table 9

Changes to tumor classification using TRF parameters only and combined TRF and DR parameters.

Similarly, Table 10 shows the changes to tumor classification using only DR parameters and the combined TRF and DR parameters. In this comparison, 35 more tumor pixels were correctly classified and 17 more were mis-classified using the combined dataset than just DR parameters, resulting in an increase of 18 more correct classifications, an increase of 5%. With a McNemar–Bowker test $p$-value of 0.02, there is a statistically significant difference between the two classification models. Both comparisons show that combining the parameters from both techniques results in improved tumor classification compared to either technique alone.

Table 10

Changes to tumor classification using TRF parameters only and combined TRF and DR parameters.

During sample collection, obtaining normal adipose tissue is much easier than obtaining normal fibroglandular tissue. As such, sampling of normal breast issue was biased toward adipose. This resulted in the sample size for adipose to be greater than fibroglandular by nearly 10-fold. In future work, it will be important to collect more fibroglandular tissue for better classification and to decrease the number of false positives.

While this technology does not yet have a tumor detection sensitivity and specificity high enough to replace microscopic pathologic examination as the gold standard in breast margin assessment, it does provide information that could give surgeons greater confidence during surgery. The system could aid in real-time decision making in the operating room, reducing surgery time, which can improve patient outcome.