**1. Introduction**

Intraoperative detection of breast malignancy margins would allow minimization of the risk of tumor recurrence in patients undergoing breast conservation surgery (BCS). Intraoperative pathological estimation can be performed through frozen section analysis and imprint cytology [1]; however, these techniques are characterized by several restrictions such as resource intensity, sampling only a small percentage of the surgical margins and limited e fficacy, especially for ductal carcinoma in situ (DCIS) [2]. Consequently, these methods have not been widely adopted [3]. Fluorescent techniques that utilize molecular contrast, potentially a ffording surgeons to visualize tumor in the cavity, are currently in the development [4,5]. Ultrasound elastography has been developed for a number of applications, and specifically for preoperative diagnosis of breast lesions [6–8]. However, its relatively low spatial resolution makes it inappropriate to use this method for intraoperative tumor margin assessment.

Optical coherence tomography (OCT) presents a very promising method for surgical tasks solving due to the clear benefits of this method such as: safety (using a near infrared light source does not risk tissue damage); accuracy (high resolution ~10–15 micron); there being no need for contrast agents; and the short duration of image attainment. OCT can be used both in the resected specimen of tumor and in the surgical cavity. OCT can be added to biopsy needle probes and can be used to guide correct sampling of tumor biopsies [9–11]. OCT is a promising method for intraoperative guidance during the resection of breast cancer and for identifying positive margins in specimens from BCS [12–15]. Recently, OCT has been proposed for intraoperative use in distinguishing tumorous and non-tumorous tissues using handheld probes [16,17]. To overcome the limited imaging depth ~2 mm typical for OCT (which usually requires su fficiently close approaching of the OCT probe towards the studied tissue), utilization of endoscopic and/or needle OCT probes are considered [11]. Moreover, it is known that structural OCT scans exhibit low contrast between tumor and uninvolved dense stromal tissue, which makes it challenging to accurately assess margin status [18]. In view of this, improvements in several aspects of OCT attract much attention, in particular, the development of functional OCT extensions based on polarization e ffects and sti ffness analysis as considered below.

Based on the birefringence of the tissue structure, polarization-sensitive (PS) OCT provides advanced imaging of collagen fibers in the breast tissue and enhances intraoperative di fferentiation of breast cancer [19–21]. Stroma state assessment is fundamentally important, because tumor collagen matrix plays a crucial role in breast cancer invasion and metastatic spreading [22]. Several studies have developed quantitative diagnosis algorithms for intraoperative assessing breast cancer margins and validated them against OCT, both alone and in combination with other modalities [23]. Cross-polarization OCT (CP-OCT) is a variant of PS OCT that allows imaging of the initial polarization state changes due to both birefringence and cross-scattering in biological tissues [24]. Only orthogonally polarized backscattered light, which is mutually coherent with the incident wave, contributes to the cross-polarized (CP) OCT image. CP-OCT is a promising method for di fferentiating tumorous from non-tumorous tissues in human breast tissues [25], human brain tissues [26,27], as well as for diagnosis of bladder cancer [28–31]. OCT can also measure attenuation, which can be helpful for improving contrast of breast imaging research [32,33].

Attention to the problem of determining tissue sti ffness (elastographic mapping) by optical coherence elastography (OCE) methods has been increasing in recent years [34–37]. Su fficiently high resolution of quantitative sti ffness maps enabled by compressional OCE opened the possibility to perform morphological segmentations of tumor tissue constituents very similar to morphological segmentation of conventional histological images [25,38,39]. In these studies of experimental tumor models on animals, this technique allowed in vivo monitoring of morphological variation in tumor tissue during tumor growth and response to therapies. In studies [25,40–42], application of compressional OCE (C-OCE) for characterization of mechanical properties of excised human breast cancer specimens was demonstrated. New possibilities for intraoperative assessment of the breast cancer borders by means of optical coherence micro-elastography (OCME) were reported in a recent study [43,44]. It has been demonstrated that OCME provides additional contrast of tumor compared to OCT alone. Additionally, the potential of OCME images for evaluation of tumor margins in specimens excised during BCS was demonstrated in [41,43]. In our previous paper it was shown that CP-OCT and C-OCE can be helpful in breast cancer margin identification, as well as for grading breast cancer subtypes [25].

For more accurate evaluation of the resection margin, it is advantageous to take into account the genetic heterogeneity of breast cancer, as well as the variety of molecular-biological and morphological features influencing prognosis of the disease course (degree of aggressiveness) and treatment selection [45,46]. Indeed, it was demonstrated that probability of tumor recurrence mainly depends on molecular-biological characteristic of the tumor [47,48], while an increase in the size of the removed tissue, free of cancer cells, is not associated with a decrease in the recurrence rate [49].

Various molecular-biological and morphological features of breast cancer are anticipated to di fferently influence the polarizing and elastic tumor and peritumoral tissue qualities. This stimulates interest in evaluation of the clinical potential of polarization-sensitive and elastographic OCT techniques for determining breast cancer subtypes (malignancy grade) and improving tumor boundary detection based on the ability of these methods to identify di fferent tumor subtypes. In this way, surgeons are expected to be provided with essential information that can improve reliability of the positive resection margin detection during BCS, at least for some breast cancer subtypes.

The goals of this research are (1) to define the visual assessment criteria required for the CP-OCT and C-OCE images in order to enable di fferentiation among various breast cancer subtypes; (2) to determine the diagnostic accuracy (sensitivity and specificity) of C-OCE in comparison with CP-OCT, for delineation of tumorous and non-tumorous breast tissues and subsequent identification of breast cancer subtypes in a blinded reader study.

#### **2. Materials and Methods**

#### *2.1. Human Breast Specimens*

This study was approved by the institutional review board of the Privolzhsky Research Medical University (Protocol #10 from 28 September 2018). All of the patients included in the study provided written informed consent. A total of 70 breast tumor tissue specimens were taken from 50 patients post partial (*n* = 35) or complete (*n* = 15) mastectomy with di fferent diagnosis (Table 1). To minimize the e ffect tissue degrading, the excised specimens were immediately placed in gauze saturated with phosphate bu ffer and closed to prevent dehydration. CP-OCT and C-OCE images of the fresh, un-fixed breast tissue were acquired within 2 h after surgical excision. The studies were done on specimens with sizes from 0.5 × 1 cm to 1 × 2 cm. Specimens were taken from central zone of tumors for diagnostics of breast cancer subtypes and in the peritumoral area for visualization of normal (non-tumorous) breast tissue. A special motorized table for convenient positioning the specimen under the OCT probe was used. The entire CP-OCT and C-OCE study of each specimen was no longer than 20 min (including preliminary sample preparation and orientation).


**Table 1.** Clinical specimens' characteristics and number of imaged specimens.

#### *2.2. Multimodal OCT Device*

This study used a common path spectral domain multimodal OCT system with a central wavelength of 1310 nm and spectral width of 100 nm, with an axial resolution of 10 μm, lateral resolution is 15 μm, a scanning depth of 2 mm in air, a scanning speed of 20,000 A-scans per second. The OCT-system acquired 3D blocks of OCT data, 2 mm in depth (in air) over 2.4 × 2.4 mm<sup>2</sup> area and 2D lateral scanning with a similar field of view were acquired in 26 s. The CP-OCT and C-OCE images were generated in real time during the acquisition process. For living tissues, real-time angiographic imaging was also possible by processing the same data [50]. For the described OCT studies, the total scanning time along a 1–2 cm trajectory on a biopsy sample was 3–5 min depending on the number of stitched images.

Structural 2D (cross-sectional images) CP-OCT images were constructed in two virtual channels, one of which was co-polarized with the incident polarization (co-polarization channel) and the other one was orthogonal (cross-polarization channel) to the incident polarization, respectively [24]. CP-OCT aims to obtain the information contained in the cross-polarization channel, which allows one to form cross-polarization images caused by birefringence of the tissue from optically anisotropic structures (evaluate the state of connective tissue component), as well as due to contribution of coherent cross-polarization backscattering on non-spherical particles and particles with dimensions much larger than the wavelength. In view of low informativity of the co-polarization images (as found in previous studies [25]), only cross-polarization images were used for diagnostic conclusions in this study.

An advanced variant of phase-sensitive compression OCE [37,39,51–55] was used to visualize inter-frame strains in the tissue and subsequently map the Young modulus. The probe was slightly pressed onto the studied sample surface, and strain distribution in the probe vicinity was reconstructed. Strain mapping was based on estimation of axial gradients of interframe phase variations of the OCT signal using the "vector" method [51,53]. The name "vector" is due to the fact that, without explicitly singling out amplitude and phase, the complex-valued OCT signals in this method are considered as vectors in the complex plane, and the phase is singled out at the very last step of the processing. Such vector representation allows one to perform flexibly-tuned amplitude-weighted averaging over the processing-window area. As a result, noisy contributions of small-amplitude pixels and, at the same time, especially strong phase errors (by ~π rad.) are very efficiently suppressed. This allows obtaining strain maps with fairly high quality even without periodic averaging (which is very important for the one-directional single-step loading of the tissue used in the described studies). In addition to the exceptionally high tolerance to various measurement noises, the vector method is very efficient computationally, so that the elastographic processing of the acquired sequence of several hundreds of OCT scans requires ~5–10 s using a "typical" PC without the necessity of GPU computations.

Another important point is that the estimated interframe phase-variation gradient is averaged over a processing window, the dimensions of which being the main factor determining the resolution of the resultant OCE scans. For a rectangular processing window with comparable axial and lateral

sizes, the resolution in strain maps is also comparable in these directions and corresponds to ~1 2 of the window size. For the described system, the window size was ~90–100 μm, which defined the strain-mapping resolution ~45–50 μm. Such a window size was chosen empirically as a compromise between worse quality of the OCE-images for smaller windows (because of insu fficiently averaged noise) and too-strong smoothing of spatial inhomogeneities for larger windows.

The next important point is quantification of the tissue Young's modulus, to enable which a reference silicone layer with preliminary calibrated sti ffness (with the Young's modulus in the range 50–100 kPa) was used as described in [42,54–56]. Of key importance in the used variant of C-OCE technique is that all OCE images are formed using a pre-selected pressure level (4 kPa in the described study) standardized over the entire image area, despite the fact that for real OCE scans, the local pressure over the lateral coordinate usually varies several times because of the non-ideally planar boundary of the sample, its mechanical inhomogeneity, etc. The pressure standardization technique is based on the usage of the reference silicone layer as a sensor of local pressure as described in detail in [55]. To synthesize such a single "standardized OCE image", a series of initial structural OCT-scans acquired during monotonic compression of the sample was first processed to obtain a series of cumulative-strain maps as described in [54,57,58]. Then vertical A-scans corresponding to the selected pressure were picked up from the initial series of cumulative-strain maps and reassembled to synthesize a single cumulative-strain image in which all A-scans now correspond to the same preselected pressure onto the tissue [55]. To be sure that the strain in silicone can be considered linearly proportional to stress (pressure), high linearity of silicone was specially verified as described in [42,54,55]. Real biological tissues usually demonstrate a pronouncedly nonlinear stress–strain law. The described C-OCE method allows one to determine this law by plotting the strain in the linear precalibrated silicone against strain in any region of interest in the tissue beneath the silicone. The elasticity of the tissue can then be estimated as the tangent Young's modulus (the slope of the stress-strain curve) corresponding to the desired pre-selected pressure. It was empirically found that for breast-cancer tissue the sought tangent modulus could be conveniently estimated as the slope of the chorde corresponding to the pressure range 4 ± 1 kPa. At lower pressures, very small strain of sti ffer regions was di fficult to estimate, whereas at higher pressures, the elasticity contrast among various tumor components became worse because of strong nonlinearity-induced sti ffening of the initially softer components of the tumor (see examples in [55]). Without such standardization the intrinsic elastic nonlinearity of breast-cancer tissues may result in uncontrollable variability of the estimated elastic modulus in di fferent measurements and even di fferent parts of the same image. This unpredictable variability may be rather significant (several times and greater) even for apparently moderate strains within a few percent [42,55]. Thus, the developed pressure-standardization procedures were critically important for enabling meaningful quantitative comparisons of elastographic data obtained from di fferent measurements.

The so-obtained OCE-images were represented in the color-coded form, such that sti ffer areas (those with weaker strain) are shown in blue, and soft areas, where deformation is greater, are shown in red.

#### *2.3. Histological Study*

After CP-OCT and C-OCE imaging of the freshly-excised sample with ye<sup>t</sup> non-modified optical and biomechanical properties, the scanned area was marked on the specimen with histological ink. Then the specimen was fixed in 10% formalin for 48 h and resectioned through the marked area, so that the plane of the histological sections coincided to the cross-sectional CP-OCT and C-OCE images. For the histological evaluation, haematoxylin and eosin (H&E) staining was used. Two independent histopathologists interpreted the histological slices photographed in transmitted light with a Leica DM2500 DFC (Leica Microsystems, Wetzlar, Germany) microscope, equipped with a digital camera. Based on histopathological analysis, all samples were classified into tumorous and non-tumorous breast tissues. The revealed histological types of breast tissue include: adipose tissue with streaks of connective tissue (number of specimens *n* = 20); fibroadenomatosis/fibroadenoma (*n* = 4); DCIS (*n* = 5); invasive lobular carcinoma (ILC) (*n* = 7); invasive ductal carcinoma (IDC) of scirrhous (*n* = 24) and solid (*n* = 10) structure (Table 1). In addition, to assess tumor aggressiveness (prognosis of the disease course) immunohistochemistry (for *n* = 46 samples) was performed, identifying five molecular subtypes of the tumors: Luminal A, Luminal B (Her2/Neo-), Her2/Neo+, Non-luminal, Triple-negative cancer (TNC). Luminal A and Luminal B (Her2/Neo-) are reported to be low-aggressive tumors characterized by predominantly favorable prognosis of disease course and treatment in comparison with Her2/Neo+, Non-luminal and TNC [24]. Furthermore, it should be noted that Luminal A and Luminal B subtypes were characterized by scirrhous architectonics, while Her2/Neo+, Non-luminal, TNC had solid structure.

The results of histopathology were compared with the corresponding CP-OCT-based and C-OCE-based findings. For the blinded reader study, all images were divided into 4 groups: adipose and normal stromal breast tissue, benign breast tissue (fibroadenoma/fibroadenomatosis), non-invasive DCIS, and images portraying cancerous features of invasive low and highly-aggressive breast cancer.

#### *2.4. Reader Analysis of CP-OCT and C-OCE Images*

A blinded reader study was performed to evaluate the statistical performance of assessing tumorous and non-tumorous breast tissues based on the CP-OCT imaging (first test) and C-OCE visualization (second test). In the study, 115 CP-OCT and 115 C-OCE images from 50 patients were interpreted by 6 readers specially trained for this OCT-based assessment (2 biologists experienced in optical imaging, but unskilled in recognizing breast cancer pathology; 2 post-graduate students of the Medical University unexperienced both in optical imaging and in recognizing breast cancer pathology; 2 surgeons skilled in detecting breast cancer pathology, but without work experience in optical imaging) who were unaware whether the image contained cancer or not. The readers were given a training set of sample CP-OCT and C-OCE images (3 images of each histological type of breast tissue).

The criteria evaluated by the readers are summarized in Tables 2 and 3. Each image group had its own set of visual criteria. The reader's goal was to distinguish between tumorous and non-tumorous breast tissues. If an image was considered to represent non-tumorous breast tissue, the reader indicated a score of "0" whether it was normal breast tissue or fibroadenoma. If the reader identified malignant lesion marks, a score from "1" to "3" was assigned to the sample depending on the estimated tumor aggressiveness. The score of "1" means that the reader thinks that the image represents non-invasive DCIS; a score of "2" means that the reader considers the cancer to be invasive, but less aggressive; a score of "3" means that the reader thinks that invasive cancer is more aggressive.

The first test was based on assessment of signal architecture in cross- polarization images (Table 2). The cross-polarization channel enables more contrast visualization of the presence and state of connective tissue in comparison with the co-polarization OCT images.

Structural features in the CP-OCT images were distinguished by the following features of the scattering intensity and lateral uniformity of the signal attenuation (Table 2):

(i) the average level of the CP-OCT signal throughout the image is visually estimated as "low" like in Figure 1(b5) or "high" for the used 0–50 dB signal range, where "low" corresponded to intensities below 25 dB, i.e., the noise range in the used scale, and "high" related to the level above 25 dB on the used scale like in Figure 1(b2);

(ii) the presence of structures with a sharp boundary between contrasting-in-brightness regions with well-circumscribed boundary architecture like in Figure 1(b3) (which was graded as "yes"/"no");

(iii) the attenuation rate as estimated by the penetration depth of the probing radiation ("high" attenuation like in Figure 1(b5) and "low" like in Figure 1(b2));

(iv) the uniformity of attenuation along the interior border of the structural CP-OCT image ("uniform" like in Figure 1(b2)/"non-uniform" like in Figure 1(b4)).


**Table 2.** Visual assessment criteria of cross-polarization optical coherence tomography (CP-OCT) images for distinguishing between non-tumorous and tumorous breast tissue.

**Table 3.** Visual assessment criteria of compressional optical coherence elastography (C-OCE) images for distinguishing between non-tumorous and tumorous breast tissue.


*n*—number of images.

The second test was based on the analysis of stiffness values distributions on C-OCE images. Stiffness maps are presented in a color palette, where hard areas (blue—above 500 kPa) indicate the presence of tumor cells, and soft regions (red—below 100 kPa) represent adipose and connective tissues. At the same time, tissues with intermediate stiffness (the predominance of orange and yellow colors corresponding to ~200–400 kPa) correspond to the presence of such degenerative changes of breast-tissue stroma as fibrosis or hyalinosis of collagen fibers. The threshold values for stiffness (Table 3) of the main types of breast-tissue components were identified as described in detail in our previous work [25], in which accurate comparison of histological and OCE images was performed.

Main and additional criteria of subsuming the images to one or another group were formulated for cross-polarization and C-OCE images, the additional criteria of visual assessment being needed for more precise differentiation among different malignant states (Tables 2 and 3).

**Figure 1.** Representative depth-wise co- and cross-polarization OCT images (**<sup>a</sup>**,**b**) of non-tumorous and tumorous breast tissue with the corresponding histology (**c**). (**a1**–**c1**) Adipose tissue with streaks of connective tissue; (**a2**–**c2**) fibroadenomatosis/fibroadenoma; (**a3**–**c3**) DCIS; (**a4**–**c4**) invasive ductal carcinoma (IDC) of scirrhous structure (low-aggressive breast cancer subtype); (**a5**–**c5**) IDC of solid structure (highly-aggressive breast cancer subtype). (**a1**–**a5**) OCT images in co-polarization channel; (**b1**–**b5**) OCT images in cross-polarization channel; (**c1**–**c5**) histological images, haematoxylin and eosin (H&E) staining. Abbreviations: A—adipose, CT—connective tissue, FA—fibroadenomatosis, ADH—atypical ductal hyperplasia, DCIS—ductal carcinoma in situ, TS—tumor stroma, TC—cluster of tumor cells.

#### *2.5. Statistical Analysis*

The results of the blinded reader study of CP-OCT and C-OCE images analysis were collected for determining the diagnostic accuracy for distinguishing: (1) non-tumorous breast tissues (*n* = 33) from tumor (*n* = 82); (2) non-invasive DCIS (*n* = 10)) from invasive breast cancer (*n* = 72); (3) low-aggressive invasive tumors (Luminal A, Luminal B (Her2/Neo-)) (*n* = 47) with favorable prognosis from highly aggressive invasive tumors (Her2/Neo+, Non-luminal, TNC) (*n* = 25) with unfavorable prognosis; (4) non-tumorous breast tissues (*n* = 33) from low-aggressive invasive tumors (*n* = 47), and (5) non-tumorous breast tissues (*n* = 33) from highly-aggressive invasive tumors (*n* = 25).

The statistical analysis was performed using Statistica 10.0 and IBM SPSS Statistics software.

The assessment of the informative value and diagnostic capabilities of the studied methods (CP-OCT and C-OCE) was carried out with an estimation of their sensitivity (Se), specificity (Sp), and diagnostic accuracy (Ac). Based on the sensitivity and specificity values, Receiver operating characteristic (ROC) curves were constructed, which show the dependence of the number of true positive rate (TP) on the number of false positive rate (FN). For quantitative characterization of the ROC curves, we evaluated the area under the ROC curve (AUC), i.e.**,** the area bounded by the ROC curve and the axis of the false positive rate [59]. The higher the AUC, the better the classifier is.

The inter-reader agreemen<sup>t</sup> was calculated using Cohen's kappa coefficient (k): k ≥ 0.81—perfect agreement; 0.61 ≤ k < 0.80—substantial agreement; k < 0.6—poor agreemen<sup>t</sup> [60].
