2.1.

Patients and Study Design

Premenopausal, nonpregnant subjects $>18$ with a clinical finding of Breast Imaging Reporting and Data System (BI-RADS) assessment category 2 and higher^13,14 were offered participation in this prospective study. All participants provided informed consent that was approved by the institutional review committee of Dankook University Hospital. All DOSI measurements took place between May 2014 and August 2015 at Dankook University Hospital, Cheonan.

All subjects underwent SOC ultrasound (US) imaging. The subjects were referred to the radiologist for US imaging because of a suspicious mammographic screening or because of a newly palpated breast mass. Patients with a finding of BI-RADS category 4 and higher underwent MRI and biopsy. If requested by patients, biopsies were also performed on BI-RADS category 3 findings. Study participants were measured by the same DOSI operator and, if applicable, at least 8 days after biopsy.

2.2.

Instrumentation

The DOSI instrument combines frequency domain photon migration (FDPM) and broadband steady-state (SS) NIR spectroscopy for quantitative, model-based broadband measurements of tissue absorption and reduced scattering spectra from 650 to 1000 nm. The full technical details of the system are described elsewhere.^15,16 Briefly, the FDPM component uses 4 laser diodes (Blue Sky Research, Sanyo, Mitsubishi) at the wavelengths 660, 680, 785, 810 nm. The breast is illuminated sequentially by each laser diode. When active, each laser diode is intensity modulated at 451 modulation frequencies swept from 50 to 500 MHz ($\sim 200\mathrm{ms}$). The diffusely scattered light that propagates through tissue is detected by an avalanche photodiode (Hamamatsu module C5658) mounted inside a temperature-controlled hand-held probe. The SS component uses a high-intensity tungsten-halogen source (Mikropack model HL-2000-HP-FHSA) to illuminate the tissue, and the diffusely scattered light that propagates through the tissue is detected by a grating-based spectrometer that collects light from 650 to 1000 nm (1024 pixels, BWTek model 611E). The FDPM and SS sources are coupled by independent optical fibers mounted into the hand-held probe. The source–detector separation was 28 mm (breast measurement) for both FDPM and SS components with fibers placed in an overlapping geometry.

FDPM and SS data are combined using a diffusion theory model to provide broadband absorption (${\mu }_a$) and reduced scattering (${\mu }_s^{\prime }$) spectra from 650 to 1000 nm. In the FDPM measurement, the phase and amplitude of the detected light are recorded as functions of source modulation frequency and fit to a diffusive model of light transport with semi-infinite boundary conditions to recover ${\mu }_a$ and ${\mu }_s^{\prime }$ at each of the four laser wavelengths. SS broadband reflectance spectra were converted into absolute absorption spectra using two steps. First, the spectral shape of the reduced scattering spectrum is assumed to follow a power law of the form ${\mu }_s^{\prime }=A·{\lambda }^{-b}$, where $A$ is the scatter amplitude and $b$ is the scatter power, or the exponent of the reduced scattering spectrum. The power-law fit to the FDPM discrete laser diode spectrum provides a scatter correction for the SS reflectance spectrum. We then fit the SS reflectance intensity at each of the laser diode wavelengths to the reflectance calculated from the FDPM-measured absolute absorption values. Thus, the SS reflectance spectrum intensity is quantitatively scaled using the FDPM discrete ${\mu }_a$ measurements. The absolute absorption spectrum is then extracted by fitting the corrected reflectance spectrum to a diffusion reflectance model.^15,16

2.3.

Measurement Procedure

DOSI measurements were performed using a standard protocol.¹⁷ With the patient supine, the DOSI probe was placed against the breast tissue and sequential measurements were taken in a rectangular grid pattern using 10-mm spacing. The opposite breast was measured for comparison. The dimensions of the grids ranged from $4\times 7$ to $9\times 9{\mathrm{cm}}^2$. The grid size was chosen to fully encompass the lesion’s anatomic extent (by US or palpation) and to capture at least $\sim 1\mathrm{cm}$ surrounding normal tissue. The grid on the contralateral breast (normal side) was mirrored from the grid on the lesion-containing breast (lesion side). An example of DOSI imaging grid locations and the resulting DOSI images are provided in Fig. 1.

Fig. 1

Benign $23\mathrm{mm}\times 11\mathrm{mm}$ fibroepithelial lesion in a 41-year-old subject. (a) DOSI measurement grids and corresponding scattering amplitude maps. (b) Mammogram in the LCC view. The arrow points at the lesion location. (c) US image of the lesion.

2.4.

Spectral Analysis

Concentrations of oxy-hemoglobin (${\mathrm{ctHbO}}_2$), deoxy-hemoglobin (ctHHb), water (${\mathrm{ctH}}_2\mathrm{O}$), and bulk lipid (ctLipid) were calculated by fitting a linear combination of their known molar extinction coefficient spectra to the tissue absorption.¹⁸ From these quantities, the total blood volume ($\mathrm{THb}=\mathrm{ctHHb}+{\mathrm{ctHbO}}_2$), percent oxygen saturation (${\mathrm{StO}}_2={\mathrm{ctHbO}}_2/\mathrm{THb}$), and tissue optical index ($\mathrm{TOI}=\mathrm{ctHHb}\times {\mathrm{ctH}}_2\mathrm{O}/\mathrm{ctLipid}$) were calculated. TOI has been shown to provide high contrast and sensitivity to chemotherapy response.⁹ Quantitative images of the local tissue concentrations, the scattering parameters, and contrast functions were formed for visualization.

Each specific tumor component (STC) spectrum was assessed by calculating the residual of the fit of the lesion absorption (corrected for normal tissue absorption) to the four principal chromophore extinction spectra. Details of this method have been reported.¹⁰ Briefly, the difference between tumor and healthy tissue absorption spectra tissue is calculated. This first differential provides a spectrum that removes spectral components common to both tissues. The difference between the aforementioned first differential and its fit to the four chromophores basis spectra forms the STC spectrum.

The STC spectrum represents the contributions of other chromophores not accounted for in the basis spectra fit (e.g., other by-products of hemoglobin such as met-hemoglobin or carboxy-hemoglobin). In addition, the STC contains information on shifts of the water and lipid peaks that correlate with different molecular states between lesions and normal tissue.

2.5.

Malignancy Index

A previously published malignancy index is used to classify lesions as malignant or benign according to their STC spectral features.¹² In brief, the algorithm identifies spectral features specific to malignant and benign lesions and maximizes the differences by weighting different wavelength regions. Weighting factors are determined by using computer processing through an iterative process for each wavelength region to determine what combination of values for each wavelength region would best separate the benign from the malignant lesions. It then translates the distance of the weighted spectra to the average weighted spectra of benign and malignant lesion into an index. The malignancy index ranges from $-1$ to 1. (A more negative value indicates a benign classification and a more positive value indicates a malignant classification.) Details about the algorithm can be found in the appendix of Kukreti et al.¹²

Kukreti et al.,¹² who introduced the malignancy index, reported that a dataset of 20 cases was sufficient to ensure reliable classification of benign and malignant lesions. Our study includes 21 lesions.

2.6.

Statistical Analysis

From the DOSI images, a region of interest (ROI) representing the lesion was defined using the US lesion size and was centered on the TOI lesion enhancement. In case the DOSI lesion enhancement was not clearly apparent, the lesion location by palpation or US (for the nonpalpable cases) was used to help localize the lesion in the TOI image. A normal tissue ROI was selected in the normal breast to mirror the lesion ROI. Only the original point measurements (not the interpolated) were included in the ROI. Lesion and normal DOSI variables were determined by calculating the mean and standard deviation (SD) of the measurement locations within the ROIs. In this work, we report the lesion to normal ratio (L/N), which accounts for the high inter-subject biological variations.^19,20 L/N was obtained in all DOSI variables, i.e., ctHHb, ${\mathrm{ctHbO}}_2$, ${\mathrm{ctH}}_2\mathrm{O}$, ctLipid, ln(A), b, THb, ${\mathrm{StO}}_2$, and TOI, except the malignancy index (normal tissue already taken into account in its calculation). Note that we use the natural logarithm of $A$ because of its wide range of values.

Box plots were used to show the distribution of the L/N DOSI variables: the box represents the 25th to 75th percentiles and median of the depicted values, the whiskers represent the SD, and the square represents the average value.

The Shapiro–Wilk test was performed to examine the normality of the distribution of all variables. The two-sample $t$-test and Wilcoxon Rank Sums test were used to compare benign and malignant lesions for the normally and nonnormally distributed variables, respectively. Statistical significance was assumed for a $p$-value $<0.05$.

Univariate logistic regression was performed to assess the relationship between the DOSI predictor and diagnosis outcome. The area under the curve (AUC) and 95% confidence interval were reported. Accuracy, sensitivity, specificity, negative predictive values (NPVs), and positive predictive values (PPVs) were also reported for an optimal threshold that would generate the minimum difference between the sensitivity and the specificity. All statistics were performed using statistical analysis system v9.4.

2.7.

Clinical Findings

Breast density is reported using the BI-RADS of breast tissue composition from the American College of Radiology.¹³ It is classified into four categories that are qualitatively based on the relative amounts of fat and dense fibroglandular tissue observed in a mammogram. Category “a” refers to almost entirely fatty tissue, category “b” to scattered fibroglandular densities that could potentially obscure a lesion, category “c” to heterogeneously dense breast type and the sensitivity of mammography may be lowered, and category “d” to extremely dense breast type that will lower the sensitivity of mammography.

Histopathology was determined from biopsy specimens for 16 out of the 21 lesions. The five remaining lesions were diagnosed as BI-RADS category 2 and 3 using US imaging with or without mammography. Per clinical guidelines, BI-RADS category 2 does not require any treatment/follow-up, and BI-RADS category 3 requires only short-interval follow-ups.¹⁴

3.1.

Patient and Tumor Characteristics

Table 1 presents the characteristics of the subjects enrolled in this study. The average (SD, range) age of subjects was 41 (9, ranging from 20 to 53). Note that the subjects with a malignant lesion were on average older than the ones with a benign lesion; 46 (4, ranging from 37 to 53) versus 35 (9, ranging from 20 to 45), respectively. Because mammography is not routinely performed on young patients, especially in case of benign findings, the breast density was available in only 13 out of the 19 subjects overall, and in 3 out of the 9 subjects with a benign lesion. As expected, high breast density (categories c and d) was observed in 11 out of the 13 subjects who underwent mammography.

Table 1

Subject characteristics.

Table 2 reports the lesion characteristics at the subject level. All malignant lesions were invasive ductal carcinoma (IDC). Fibroadenoma accounted for $\sim 64\%$ of the benign lesions. Histopathology of all malignant lesions was determined by biopsy and confirmed at surgery. Seven out of eleven benign lesions were diagnosed by biopsy. One benign lesion was diagnosed as fibroadenoma using US imaging, and required 6-month follow-up appointments. One benign lesion that had been diagnosed as fibroadenoma two years prior to the study did not show any change in size during the 6-month follow-up visits. One benign lesion was diagnosed as an oil cyst using US imaging. One benign lesion was diagnosed as fibro-cystic change using US imaging and mammography.

Table 2

Lesion characteristics.

The average (SD, range) lesion size measured by US imaging (greatest dimension) was 2.0 cm (1.3 cm, 0.7 to 5.6 cm). The average (SD, range) lesion depth measured by US imaging was 0.8 cm (0.4 cm, 0.1 to 1.6 cm). Notably, while the malignant and benign lesions were located at similar depth in the breast, their average sizes differed substantially: malignant lesions were on average larger than the benign lesions, with average (SD, range) sizes of 2.6 cm (1.5 cm, 1.2 to 5.6 cm) and 1.7 cm (0.8 cm, 0.7 to 2.8 cm), respectively. All malignant masses were found to be palpable, while only 5 out of 11 benign masses could be palpated.

DOSI enhancement of the lesion was not clearly apparent for 8 out of the 21 cases. For two benign lesions, the DOSI enhancement was similar to normal tissue. For six lesions, five benign and one malignant, the enhancement was partially mixed with the areolar enhancement. For these eight cases, the lesion location by palpation or US (for the nonpalpable cases) was used to help localize the lesion in the TOI image to calculate the lesion ROI.

3.2.

DOSI and SOC Imaging

Figures 1 and 2 present two typical examples of benign and malignant lesions measured using DOSI and using SOC imaging.

Fig. 2

Malignant $14\mathrm{mm}\times 14\mathrm{mm}$ IDC in a 47-year-old subject. (a) DOSI measurement grids and corresponding TOI maps. (b) Mammogram in the LMLO view. The arrows point at the lesion location. (c) US image of the lesion. (d) Axial cross section of the MRI image of the chest.

3.3.

Comparison of DOSI Parameters in Malignant and Benign Lesions

Figure 3 compares the averaged STC spectra for the benign lesions, malignant lesions, and normal tissue on the contralateral side. As expected, due to the specificity of the STC to tumors, we observe little to no spectral features in the STC spectra of the normal tissue. For malignant lesions, the spectral content shows greater magnitude than for the benign lesions. Distinctive spectral differences also exist between the STC spectra of the benign and malignant lesions.

Fig. 3

Average STC spectra for 10 malignant lesions (solid black), 11 benign lesions (dashed black), and 21 contralateral normal tissue regions (solid gray). Error bars show the distribution of the population.

Figure 4 and Table 3 compare the distribution of lesion to normal (L/N) in all DOSI parameters between the benign and malignant cases. Based on the Shapiro–Wilk test, only ${\mathrm{StO}}_2$ and the malignancy index were found to be normally distributed (data not shown). As a result, the two-sample $t$-test was applied to ${\mathrm{StO}}_2$ and the malignancy index, and the Wilcoxon Rank Sums test to the rest of the variables to assess the statistical difference between benign and malignant lesions. As observed in Fig. 4 and Table 3, the malignant group exhibited on average greater L/N values than the benign group for all DOSI parameters except for the lipid. These differences were found to be statistically significant in all parameters except for the lipid and ${\mathrm{StO}}_2$ ($p$-value provided in Fig. 4 and Table 3). A complete list of absolute DOSI parameters for benign and malignant lesions is provided in Table 4. Malignant lesions exhibited greater ctTHb, ${\mathrm{ctHbO}}_2$, ${\mathrm{ctH}}_2\mathrm{O}$, TOI, ctHHb content than benign lesions; however, a detailed statistical comparison was not performed.

Fig. 4

Distribution of L/N in the groups of benign (plain) and malignant (striped) lesions in (a) scattering power $b$, ln(A), ${\mathrm{HbO}}_2$, HHb, water, lipid, THb, ${\mathrm{StO}}_2$, (b) TOI, and (c) malignancy index. Mean value in each group is indicated with a square. Boxes show the median and the 25th and 75th percentiles. Whiskers show the SD. The blue star indicates a statistically significant difference between the two groups (Wilcoxon Rank Sums test, $p$-value provided).

Table 3

Mean, SD, and range of the L/N optical parameters in the malignancy and benign groups.

Table 4

Mean, SD, and range of the optical parameters in the malignancy and benign lesions.

3.4.

Performance of DOSI for Differential Diagnosis

Tables 5 and 6 summarize the power, assessed by area under the ROC curve (AUC), of L/N to predict a malignant lesion from a cohort of malignant and benign lesions in all DOSI parameters. The total hemoglobin and the malignancy index were the two best unique predictors among the DOSI parameters with an AUC of 0.96 (95% CI 0.86 to 1.00) and 0.99 (95% CI 0.97 to 1.00), respectively. At the specific threshold specified in Table 6, this resulted in accuracy, sensitivity, specificity, PPV, and NPV of more than 90% for both the malignancy index and ${\mathrm{THb}}_{\mathrm{L}/\mathrm{N}}$. Overall, all DOSI parameters except ${\mathrm{stO}}_{2-\mathrm{L}/\mathrm{N}}$ exhibited statistically significant prediction of malignancy, with $\mathrm{AUCs}>0.77$.

Table 5

Performance of the DOSI parameters for differential diagnosis: AUC and 95% confidence interval for all DOSI parameters.

Table 6

Performance of the DOSI parameters for differential diagnosis (follow-up of Table 4): accuracy, sensitivity, specificity, PPV, and NPV at the specified threshold for all DOSI parameters.

3.5.

Influence of Discrepant Averaged Lesion Size Between the Two Groups on the Differential Diagnosis Power of DOSI

As shown in Table 2, the average size of malignant lesions was much larger than that of the benign lesions (2.6 versus 1.7 cm, respectively). Furthermore, while all malignant lesions were palpable, only 5 out of 11 benign lesions were also palpable.

In order to assess the impact of the difference in the average lesion size between the two groups to the analysis, we evaluated the power of DOSI to differentiate benign from malignant lesions in a subset of lesions with similar size and depth. As depicted in Fig. 5, 12 patients (6 benign and 6 malignant) were selected based on their similar tumor size and depth. The resulting average (SD, range) lesion sizes were 1.8 cm (0.6 cm, 1.2 to 2.8 cm) and 2.3 cm (0.5 cm, 1.5 to 2.8 cm) for the malignant and benign lesions, respectively. Out of the six benign lesions, five were palpable. Lesions from both groups were located at similar depths (SD, range) into the breast tissue: 0.7 cm (0.3 cm, 0.2 to 1 cm) and 0.6 cm (0.3 cm, 0.3 to 0.9 cm) for the malignant and benign lesions, respectively.

Fig. 5

Subsets of patients used in testing the impact of the average lesion size on the different diagnosis power of DOSI.

In this subset of data, we observed again greater lesion to normal values in the malignant group compared to the benign group for all DOSI parameters except for the lipid and oxygen saturation (data not shown). These differences were found to be statistically significant in THb, ${\mathrm{ctH}}_2\mathrm{O}$, TOI, and malignancy index ($p\text{-}\text{value}<0.022$), and to be near-significant in ctHHb and ${\mathrm{ctHbO}}_2$ ($p\text{-}\text{value}=0.054$). The performance of DOSI for differential diagnosis in this subset of patients is presented in Table 7. Comparable with the findings using the whole dataset, the malignancy index and total hemoglobin were the best unique predictor of malignancy, with AUCs of 1.00 (95% CI 1.00 to 1.00) and 0.94 (95% CI 0.82 to 1.00), respectively.

Table 7

Performance of the DOSI parameters for differential diagnosis in the subset of patients: AUC, 95% confidence interval, and p-value for all DOSI parameters.