1.

Introduction

Breast cancer is the most common cancer found in women in the United States, with an estimate of over 180,000 new cases every year. It is also the second leading cause of cancer deaths in women, after lung cancer, and is responsible for the deaths of over 40,000 women every year in the U.S.¹ The most common breast malignancy is invasive ductal carcinoma (IDC), which accounts for around 75 to 80% of invasive breast cancers, while invasive lobular carcinoma (ILC) accounts for most other invasive cases. As with all malignancies, early detection of breast cancer is the only way to effectively manage patients who suffer from this disease. According to the American Cancer Society (ACS),² the $5\text{-}\text{year}$ survival rates for persons with breast cancer that are appropriately treated are as follows: 100% for stage 0, 100% for stage I, 92% for stage IIA, 81% for stage IIB, 67% for stage IIIA, 54% for stage IIIB, and 20% for stage IV. This indicates that improved diagnosis methods need to be developed to detect this cancer in its early stage when it is treatable and its survival rates are high.

Mammography is currently the standard screening tool for breast cancer. When suspicious masses are found during the mammographic screening, various other tests are performed to confirm and stage the disease. Among these tests, biopsy has proven to be the best way to determine whether a suspicious area seen in an image is in fact cancer. Breast biopsies provide very useful diagnostic information and can be comfortably performed with intravenous sedation and local anesthesia. However, while the biopsy of relatively large palpable masses usually has a high diagnostic yield, up to 99% (Ref. 3), the biopsy of smaller infiltrating masses has a much lower diagnostic yield (below 60% in some studies).^{4, 5, 6, 7, 8, 9, 10} Without a priori knowledge of the 3-D location of small cancer masses, it is unlikely that a biopsy protocol will yield consistently high cancer detection rates. This is the primary reason why biopsy has relatively high rates of false-negative diagnoses when no guidance modalities are used. Since the majority of lesions for which women undergo biopsy prove to be benign and many women have multiple biopsies during their lifetimes, very reliable and less invasive techniques are required. Currently, three types of biopsies are used on a large scale: fine needle aspiration biopsy (FNAB), core needle biopsy (CNB), and vacuum biopsy (VB).

FNAB is the least invasive and best tolerated procedure, typically using a 23-gauge needle or smaller,¹¹ and it is preferred by many patients because it does not produce discomfort or bleeding. However, its diagnostic yield largely depends on the clinician skills.^{10, 12, 13, 14} FNAB is a cost-effective and rapid procedure, and the results are rapidly available for cytopathologic analysis. Therefore, the patient usually has a diagnostic by the end of the visit. The number of annual FNAB procedures varies from one clinic to another within a range of a few hundred to thousands,¹⁵ but even for small clinics, it is usually over one hundred per year.

CNB is the most used technique and has high reliability because it allows for collection of larger size tissue samples (up to $1.5\mathrm{mm}$ in diameter and $1\text{to}2\mathrm{cm}$ in length) for histological analysis. However, CNB is often associated with deleterious side-effects, including serious discomfort and bleeding, and potential tissue morbidity. Also, CNB requires ultrasound or computed tomography (CT) needle guidance to sample nonpalpable masses and to avoid perforating major vessels.

VB has recently gained acceptance by clinicians because it provides more reliable results than FNAB or CNB.^{16, 17} However, it is relatively expensive and produces even greater discomfort than CNB.

Currently, most of the FNAB interventions are done without any guidance modality. However, due to the inability to identify tissue type by manual palpation and the challenges of positioning the needle tip within a viable tumor, which may be admixed with normal, reactive, and necrotic tissue, nondiagnostic aspirates occur in about 20% of aspirates and in 5 to 15% of patients.³ Therefore, a relatively simple but efficient method for FNAB guidance would substantially increase the diagnostic yield of this simple, minimally invasive, and affordable procedure. Proper needle placement could substantially reduce the number of the nondiagnostics aspirates and improve the sensitivity and specificity of the procedure. Ultrasound and stereotactic CT guidance of all three types of biopsy procedures have produced enhanced outcomes.^{18, 19, 20} However, ultrasound and stereotactic CT guidance is not always available and when it is available, it substantially increases the overall cost of the biopsy. Therefore, simpler and less expensive guidance methods are desirable.

Optical methods have been developed for many years to improve biopsy outcome. They can detect tissue abnormalities with relatively good accuracy, and therefore they offer a viable alternative for biopsy guidance. Among various techniques developed to date, spectroscopic-based methods have shown real promise for tissue-type discrimination. For example, several diagnostic studies have found significant differences in both the emission and excitation spectra from normal, benign, and/or malignant breast tissues. Alfano ^{21, 22} have first showed differences between the fluorescence emission spectra of normal and malignant breast tissues. Yang ^{23, 24} have used the emission spectra within the $300\text{to}400\text{-}\mathrm{nm}$ spectral region to discriminate between malignant and fibrous samples. They found 93% sensitivity and 95% specificity, but results were worse for discriminating normal fatty and malignant tissues. Gupta ²⁵ have measured emission spectra when normal, benign (fibroadenoma), and malignant (IDC) breast tissue samples were excited with $337\mathrm{nm}$ . Using the integrated emission intensity from the $337\text{-}\mathrm{nm}$ excitation, malignant tissues were separated in a binary fashion from both benign and normal with a specificity of 98%. Diffuse reflectance methods have also shown promise for use in tissue discrimination. Several studies have determined that diffuse reflectance spectroscopy can detect changes in scattering and absorption due to malignancy-associated alterations in levels and organization of hemoglobin, $\beta$ -carotene, DNA, and other proteins.^{21, 22, 23, 25, 26, 27, 28, 29} Bigio ²⁶ have used in vivo measurements to distinguish between malignant and normal tissue with sensitivities up to 69% and specificities up to 93%. Ramanujam³⁰ has combined reflectance and fluorescence measurements but has found no significant improvement in diagnostic performance using this multimodal approach. However, neither of the aforementioned reflectance technologies can probe the tissue in depth over a distance of at least $1\text{to}2\mathrm{mm}$ , and therefore their role is mostly limited to epithelial malignancy biopsy guidance (for example, for colon polyps, esophageal, oral, or cervical cancer). Another limitation of most of these techniques is that they cannot be performed through the lumen of a fine gauge needle. Only optical reflectance and the reduced scattering coefficient have been investigated using a needle-like probe for tissue characterization.³¹

More recently, optical coherence tomography (OCT) and low-coherence interferometry (LCI), the nonscanning, nonimaging version of OCT, have been applied to tissue discrimination toward optical biopsy^{32, 33, 34, 35, 36} and image-guided surgery of breast cancer.³⁷ OCT is an optical ranging technique that is capable of imaging depth-resolved (axial, $z$ ) tissue structure, birefringence, flow (Doppler shift), and spectra at a resolution of several microns. The tissue probing depth with this technology is on the order of $2\text{to}3\mathrm{mm}$ , which is almost one order of magnitude higher than the probing depth of the spectroscopic approaches. Besides this advantage, OCT systems can be constructed using fiber-optic components, and therefore OCT probes can easily fit within the bore of a fine gauge needle (for both scanning and nonscanning modes), allowing diagnostic information to be obtained directly from the FNAB site. Such very small fiber-optic-based probes have numerous clinical diagnostic and therapeutic applications.

Automated interpretation of OCT findings is, nonetheless, a very challenging issue. Previous studies^{32, 33, 34, 35, 36} suggested that this technology has the potential to substantially increase the diagnostic yield of the FNAB procedures. However, until now, only the differentiation between adipose and tumor has been demonstrated on both OCT and LCI scans.^{33, 34, 35, 36} A previous study³² demonstrated the possibility of differentiating between tumor, adipose, and stroma (connective tissue) using elaborated algorithms illustrated on a very limited number of samples and without axial discrimination. The capability of differentiating between the multiple tissue types (adipose, fibrotic, tumor, necrotic, etc.) that could be present within the same OCT or LCI scan will add more value to an optical guidance tool.

In this paper, we demonstrate an advanced algorithm for automated differentiation of the three major tissue constituents: adipose, fibrous, and tumor, usually found admixed in suspicious breast masses. The algorithm was tested ex vivo on 137 samples of human breast tissue and provides spatial discrimination of the tissue both lateral and axial. With this algorithm, the pathologist/clinician performing the biopsy will be able to more precisely determine what tissue type is present at the tip of the needle before performing the biopsy. Therefore, this algorithm could help substantially decrease the number of nondiagnostic aspirates and increase the overall biopsy yield.

2.

Methods

2.1.

Instrumentation and Measurement Protocol

An ex vivo study on excised tissue specimens was conducted to test the capability of the OCT technology for tissue differentiation. The main objectives of this study were to identify characteristic features of each of the tissue types (fibrous, adipose, tumor), develop quantitative metrics for tissue differentiation using a training set of tissue specimens, and test these metrics on a validation set of tissue specimens.

OCT measurements were performed on over 150 fresh tissue samples from patients with breast cancer surgery (lumpectomy and mastectomy). A $1310\text{-}\mathrm{nm}$ spectral-domain OCT (SD-OCT) system, presented elsewhere,³⁸ was used for this study. This system provided an axial resolution of $10\mu \mathrm{m}$ , a lateral resolution of $25\mu \mathrm{m}$ , and an imaging range of about $2.2\mathrm{mm}$ at a line rate of $5.12\mathrm{kHz}$ . The SD-LCI/OCT system can accommodate both scanning and nonscanning probes, and therefore it can work in both LCI and OCT modes. A bench-top OCT probe was used in this study employing a free-space scanning mechanism in the sample arm. An imaging rate of $5\text{frames}∕\mathrm{s}$ of $512\times 1024\text{pixels}$ /frame was achieved with this system and was limited by the relatively low reading rate of the InGaAs line detector (SU512LX, Sensors Unlimited, Inc., Princeton, New Jersey). However, we are currently upgrading this system, and the new SU1024-LDH camera will allow for a frame rate of $45\text{frames}∕\mathrm{s}$ .

The breast tissue samples were obtained from the Pathology Department, Massachusetts General Hospital (MGH), and National Disease Research Interchange (NDRI). No information about tissue donors was provided. Tissue procurement, handling, and data collection were performed according to an MGH-approved Institutional Review Board protocol (2002P000487 from 09/14/2007) and NDRI protocol (DIFN1-001-005 from 03/05/2007). Tissue samples from NDRI were shipped overnight within a few hours from excision. The tissue was kept in saline and shipped in a box with ice bags and arrived in very good pathological condition.

Our tissue measurement protocol consisted of OCT B-scans over small areas of each sample ( $3\text{-}\mathrm{mm}$ lateral scans). Each tissue sample was kept hydrated in saline solution at $37°\mathrm{C}$ during the measurements. After completion of OCT measurements, each tissue sample was marked with india ink on the OCT imaging locations (usually 3 to 5 locations on each sample) and fixed with formalin (10% formalin in PBS). Histologic preparation of each tissue specimen was then performed at the MGH histology department.

Typical OCT images for each tissue type are shown in Fig. 3. The OCT image was cropped to $1.5\mathrm{mm}$ depth to keep only the tissue part of the image. Representative depth reflectivity profiles (A-scans) are shown in the first row of Fig. 1. A clear difference can be observed between the adipose tissue and the other two tissue types (fibrous and tumor). However, this difference is less significant between fibrous and tumor tissue. In many cases, tissue differentiation is difficult, especially within breast masses that consist of admixed tissue types, when more than one tissue type is present within the same reflectivity profile. Therefore, a set of key metrics (signal slope and variance, mean spatial frequency of the intensity peaks, mean area of the power spectrum peaks, etc.) was developed to find specific characteristics for each tissue type.

Fig. 1

Graphical illustration of the main steps in the signal processing sequence. First column (position 1 in Fig. 3, shown laser)—adipose tissue; second column (position 2 in Fig. 3)—fibrous and adipose tissue; third column (position 3 in Fig. 3)—tumor tissue. First row—unprocessed depth profile (continuous) and smoothed version (dotted line); second row—smoothed profile with linear fit for each region; third row—depth profile variation around the linear fit; fourth row—normalized power spectra for each linear fit window.

Fig. 3

Representative examples of OCT diagnosis on breast tissue specimens. (a) adipose; (b) admixed fibroadipose; (c) fibroglandular; (d) tumor.

2.3.

Decision Algorithm

As a result of the analysis presented earlier, eight parameters (Slope, Std, MeanPeakDistance, StdPeakDistance, MeanFrequency, StdFrequency, PeakNr, PeakArea) were calculated and assigned to each pixel in the OCT images for each tissue specimen used in the training set. The training set is selected based on the histological diagnosis provided by an experienced pathologist.

Mean values ${\overline{x}}_i$ of each parameter are calculated for the three tissue types: adipose, fibrous, and tumor, $i=\mathrm{1,2,3}$ . ${\overline{x}}_i$ is a column vector made of the eight means. Covariance matrices are also calculated for each tissue type accounting for all eight parameters³⁹:

Eq. 1

$S_i=\frac{1}{n_i}\sum _{j=1}^{n_i}(x_{i,j}-\overline{x_i}){(x_{i,j}-\overline{x_i})}^T,$

where $n_i$ is the number of elements in each tissue class within the training set, and the superscript $T$ indicates matrix transpose. The mean values for each parameter used in our algorithm, corresponding to the three tissue types in the training set, are listed in Table 1 .

Table 1

Mean values for the eight parameters.

	Slope (a.u.)	Std (a.u.)	PeakArea (a.u.)	PeakNr	MeanPeak Distance (μm)	StdPeak Distance (μm)	Mean Frequency (105m−1)	Std Frequency (105m−1)
Adipose	$-1.027$	128.461	7.977	1.875	154.8	17.7418	7.795	8.374
Fibrous	$-3.031$	124.675	8.611	1.924	94.6	23.4178	7.581	7.279
Tumor	$-3.308$	70.78	15.251	2.761	55.9	13.76	9.348	7.542

For each sample to be diagnosed, the mean values and the covariant matrices are used to calculate a quadratic discrimination score³⁹:

Eq. 2

$d_i^Q=-\frac{1}{2}\ln \mid S_i\mid -\frac{1}{2}{(x-\overline{x_i})}^T{S}_i^{-1}(x-\overline{x_i}),$

where ∣.∣ indicates the matrix determinant, $S_i^{-1}$ is the inverse matrix of $S_i$ , and $x$ is the column vector made of the eight calculated parameters for that sample. Three quadratic discrimination scores are obtained for each pixel corresponding to the three tissue classes, and the maximum score is selected to assign each pixel of the image to the correct tissue type. The quadratic discrimination score is the logarithm of the probability that the tissue at that pixel belongs to a tissue class and the maximum probability is used for tissue assignment.

Figure 2 shows a scatter plot illustrating the clustering of the three main tissue types (adipose—green, fibrous—blue, tumor—red) and their projections on the $x$ , $y$ , and $z$ planes for only three parameters: Slope, Std, and PeakArea. The points represented here are for each pixel of the OCT images corresponding to the training set of tissue samples. With only three parameters, there is significant overlap for the three tissue types. One can notice however, that there is a decent degree of separation among the three tissue types in the Slope–Std projection plane. The adipose tissue is also well separated from fibrous and tumor in the Slope–PeakArea projection plane, while the tumor tissue is better separated from adipose and fibrous tissue in the Std–PeakArea projection plane. Using more parameters in a multidimensional space is therefore expected to produce better clustering of the three tissue types with much less overlap.

Fig. 2

Scatter plot illustrating the clustering of the three main tissue types (adipose—green, fibrous—blue, tumor—red) and their projections on the $x$ , $y$ , and $z$ planes for three parameters: Slope, Std, and PeakArea.

Multiple depth reflectivity profiles are acquired and processed in each OCT frame in either scanning or nonscanning mode. The result of the algorithm calculation is a numerical value (1 to 3) representing a tissue type. A specific color corresponding to each numerical value is attributed to every pixel in the frame: light $\text{blue}=1$ to adipose tissue, $\text{yellow}=2$ to fibrous and fibroglandular tissue, and $\text{red}=3$ to tumor tissue. Dark $\text{blue}=0$ corresponds to pixels that were masked due to low signal value. Averaging schemes, user selectable over a window of $20\times 20\text{pixels}\text{to}50\times 50\text{pixels}$ , are applied before displaying the results. Each pixel is finally assigned the dominant tissue type within the averaging window. The algorithm also calculates the percentage of each tissue type present in each frame.

We note that no image processing schemes were used here. The described algorithm was designed and implemented for A-line processing and is capable of identifying different tissue types within each A-line whether the A-lines were acquired with or without scanning (OCT or LCI mode). Averaging was done only at the end before displaying the results. It was performed to remove variability between neighboring pixels and to ensure a locally smooth display map of the tissue assignment. Alternatively, multiple reflectivity profiles can be averaged first and then only the average profile is processed. This modality is applicable to the LCI mode, where multiple A-lines are collected from the same tissue location. This can speed up processing, since 50 to 100 A-lines are sufficient to get a correct estimation, and the acquisition and processing of a small number of A-lines is very fast.

3.

Results and Discussions

Selected OCT frames for several cases of single tissue type [(A), (C), (D)], as well as of admixed tissues (B) are presented in Fig. 3 . The diagnostic maps reflect the spatial distribution of each tissue type. It can be observed that each tissue type is well recovered. In Fig. 3, the adipose (A) and fibroglandular (C) tissue types can be reasonably well differentiated by a trained OCT reader by examining the OCT image only. However, it is very difficult to diagnose the relatively small differences between the fibrous (C) and tumor tissue (D) based on the OCT appearance only. For these cases, the consensus among OCT readers is relatively low. Our algorithm, however, correlates well with the histology findings.

OCT measurements were performed on 152 tissue samples to test the capability of our algorithm for tissue differentiation. Each measurement site was marked with ink, and histology was performed to correlate OCT measurements with histology findings. However, for 15 samples the technician could not find the ink marking when slicing the tissue, and therefore these samples were removed from the study. Of the remaining 137 samples, 48 were assigned to a training set and 89 to a validation set. The training set allocation was based on pathologist recommendation. These samples showed the best representation of the three tissue types: adipose, fibrous, and tumor. A correlation of over 95% was found when the algorithm was retrospectively applied to the training set. The pathologist, blind to the algorithm findings, also performed histology readings on the validation set. The trained algorithm was then applied to the validation set, and algorithm findings were correlated with histology readings. 93% of the adipose samples were correctly diagnosed, while fibrous and tumor tissues were correctly identified in 75.5% and 88% of the samples, respectively. The same set of samples was measured in an LCI (nonscanning) configuration following the same protocol.³⁸ However, the OCT mode seems to provide better results. We attribute this to the larger tissue volume sampled in the OCT mode.

Our primary interest in this study was to train the algorithm to distinguish between normal (adipose, fibroadipose, or fibroglandular) and abnormal tissue (tumor or tumor admixed with normal tissue), and also to preferentially recognize adipose tissue, which usually creates nondiagnostic aspirates (fatty fluid or fatty cells). The sensitivity and the specificity of the algorithm findings were calculated as:

The results of our study are summarized in Table 2 . We note here that the results were obtained by processing multiple A-lines in each specimen generated by scanning a relatively large tissue section. Sensitivity and specificity of 0.88 were found. These are very good values considering the fact that the tissue differentiation parameters were based on a relatively small number of samples in the training set. The algorithm can be further improved by using a larger training set and by applying a weighting function to each key parameter used in the algorithm.³⁹ Some of the parameters may provide redundant information, and their uniqueness is still under investigation. However, given the large variability in biological tissue, they seem to behave differently across a large number of tissue samples. The algorithm can also be trained to minimize the $FN$ results (increase the sensitivity of the findings) by using a log-linear modeling to determine weighting factors for each tissue feature.³⁹ This will indicate how strongly each feature correlates to the histopathologic diagnosis. The classification approach used here follows from the assumption that each class is described by a multivariate normal distribution.³⁷ Determining more accurate probability distribution functions for each class based on a larger training set might improve the algorithm performance.

Table 2

The results of the automated algorithm.

The current version of the algorithm implemented in MATLAB (The MathWorks, Inc., Natick, Massachusetts) is not fully optimized yet. Post-processing on a laptop with a $2.0\text{-}\mathrm{GHz}$ dual-core processor currently takes $56\mathrm{s}$ for an OCT image of 1000 A-lines and 256 depth points. The processing time could be significantly improved with parallel processing of A-scans, using a faster processor, reducing the number of A-scans, and algorithm optimization. The algorithm can also be implemented in hardware for real-time processing suitable for clinical applications.⁴⁰ The study presented here was applied to OCT images. Alternatively, since the processing algorithm was intentionally designed for line processing and does not use any region-based analysis techniques (e.g., texture or kernel processing schemes), it is equally well suited for nonscanning protocols in FNAB applications where a single fiber is inserted through a biopsy needle and data from a fixed location is taken at a time.^{34, 38} In the nonscanning case (LCI mode), several A-lines are acquired, individually processed, and averaged in the end, and their processing becomes faster than of 1000 A-lines OCT images as described here. Alternatively, multiple-depth reflectivity profiles could be averaged in the nonscanning case to reduce noise, and only the average profile could be processed with our algorithm. Therefore, even the current version of the algorithm might become suitable for real-time guidance of needle biopsy.

4.

Conclusions

A novel algorithm for automated classification of breast tissue types based on OCT data was demonstrated. The algorithm was able to successfully differentiate three breast tissue types (adipose, fibrous, and tumor) providing both lateral and depth discrimination. Identification of healthy, normal versus diseased, cancerous tissue was done with a sensitivity and specificity of 0.88. An increase in both sensitivity and specificity might be possible by further refining the algorithm (as described earlier).

The algorithm was preliminarily tested on OCT images. Alternatively, since it is based on the processing of individual A-lines (does not use image features for tissue classifications), it is well suited for automatic interpretation of LCI data. This enables the use of simpler probes for biopsy guidance, consisting of a bare fiber inserted through a biopsy needle.^{34, 36, 38} Besides the use of a simpler probe, the LCI mode allows for processing of a reduced number of A-lines than in the OCT mode as well as for averaging first multiple A-lines and processing only the average profile. This mode makes the current algorithm faster and suitable for real-time tissue classification and therefore for guidance of the biopsy needle by providing the physician relevant information about the type of tissue present at the tip of the needle. This could positively impact the diagnostic yield of the FNAB procedures. Since FNAB is much less expensive and faster than CNB and VB, comparable yield on palpable masses could favor LCI-guided FNAB to become the preferred diagnostic modality.

The application of the algorithm to automated interpretation of OCT data could have a significant impact on clinical translation of OCT. Even the reported level of accuracy in differentiating tissue types could increase the yield of the biopsy procedure if OCT is used as a guidance tool. With this simple technology, the pathologist or clinician performing the biopsy will be able to guide the needle or the biopsy forceps to the most representative diagnostic area of the mass based on the instrument’s ability to determine the tissue type in real time. This will avoid unnecessary biopsy and increase the effectiveness of the procedure.