1.1.

Breast Cancer

Breast cancer remains one of the leading causes of death among women, claiming a projected 40 410 lives in the United States in 2005.¹ Early diagnosis remains a key emphasis in lowering this mortality rate. Accordingly, new optical screening tools are being developed to improve detection.^{2, 3, 4} Additional work is also necessary, however, to improve the accuracy of the approximately one million biopsies that are performed in the United States each year. Clinical biopsy localization techniques, such as stereotaxis and sonography, yield misdiagnosis in over 10% of nonpalpable lesions examined.^{5, 6, 7, 8} Improvements in the tools used for biopsy guidance will reduce this rate, thereby potentially reducing treatment costs and improving patient care. Additionally, open surgical procedures may be optimized by improving tumor margin identification, a process that often contains a great deal of ambiguity, especially in the case of some in situ carcinomas, which are not visible on gross examination.⁹

The adult female human breast consists of two main tissue types. Adipose tissue stores fat in adipocytes, large lipid-filled cells, and makes up the vast majority of the breast, while glandular structures allow for milk production and transport.¹⁰ The glandular tissue comprises a network of lobules and branching epithelial ducts, which connect the network that allows for milk delivery. In addition, systems of vascular, connective, and neural tissues reside among the glandular and fat structures. The connective tissues are often referred to as the breast stroma.

Neoplastic growth typically originates in the glandular structures, most commonly in the ducts where it presents itself as a dense cluster of epithelial cells.¹¹ After identification of a mass via palpation or mammography, typical treatment standards call for a core-needle or fine-needle biopsy and possible subsequent open surgical removal of any remaining diseased tissue.¹² Pathological diagnosis of removed tissue is based on a variety of factors including the cell types present, their spatial arrangement, and the presence of nuclear pleomorphism, among many others.¹³ Staging of the disease state is also performed in order to assess the need for treatment of metastatic disease. The potential roles for optical imaging in this treatment regimen, beyond the rapidly developing field of optical mammography,^{2, 3, 4} are as a tool for the detection of lymph node metastasis,¹⁴ a guide for needle biopsy procedures,¹⁵ an aid for the identification of margins in open surgical procedures, and, ultimately, a diagnostic tool for the evaluation of disease states in vivo.

1.2.

Optical Coherence Tomography

Optical coherence tomography (OCT) is an interferometric optical imaging modality that provides both high spatial resolution and cross-sectional imaging capability. The sample being examined is exposed to low-coherence near-infrared (NIR) light and reflections are detected using a coherence-gating scheme that allows for depthwise spatial mapping.¹⁶ Coherent detection is typically achieved using a Michelson interferometer, which combines light returning from the sample with that from a scanning delay line. The interfered light is incident upon a photodetector, which converts the optical response into an electrical signal that can be stored and analyzed using a computer system.

OCT systems, by employing broad bandwidth sources, have achieved axial resolutions below $2\mu \mathrm{m}$ , making whole cell imaging feasible,¹⁷ and approaching that of conventional histology.¹⁸ In scattering tissues, penetration depths of up to $2\text{to}3\mathrm{mm}$ , depending on tissue type, can be realized by exploiting the naturally occurring “biological window,” the spectral region between 700 and $1300\mathrm{nm}$ wherein scattering events dominate the attenuation process in tissue.^{19, 20} Because this depth is shallow relative to whole-organ or whole-body imaging modalities, catheter-, endoscope-, and needle-based OCT beam-delivery instruments are often employed for medical applications.^{17, 21, 22}

Recently, OCT imaging of mammary tissue has been demonstrated.²³ Images of carcinogen-induced rat tumor models showed that the OCT response from tumor and adipose tissues are clearly distinguishable, revealing the possibility of OCT-based techniques for the surgical identification of tumor margins and the guidance of biopsy procedures.²³ In order for these techniques to meet with widespread clinical use, however, there must be low operator training requirements. As in other clinical imaging modalities,^{24, 25} computer-aided tissue identification techniques are expected to assist in this transition. While computer-aided OCT tissue identification based on speckle analysis has been investigated in the past,²⁶ we present techniques specifically tailored to the unique structural features of breast tissue in OCT images.

1.3.

Computational Tissue Identification Techniques

Computational techniques already enjoy widespread clinical use for screening, laboratory analysis, and image evaluation. Modern analysis of cervical Pap smears, for example, is performed by computerized microscope systems that locate abnormal cells and tag them for review by a cytotechnologist or cytopathologist. These systems may employ geometrical classification, density analysis, or neural network processing to identify suspicious areas in digitized microscope images.^{27, 28, 29, 30} Computational techniques are also heavily used in digital radiology for object size measurement, boundary determination, and the automated identification of features in a variety of imaging modalities.^{31, 32} For example, digital analysis techniques have been heavily studied as an automated means by which to identify masses and microcalcifications in digital mammography images, potentially improving advanced identification of suspicious lesions.^{33, 34, 35}

2.1.

Tissue Properties

Prior to the development of effective algorithms, it is necessary to have an understanding of the OCT image properties from each tissue type. In order to develop this knowledge, human tissue was acquired from three breast cancer patients under protocols approved by the institutional review boards of the University of Illinois at Urbana-Champaign and the Carle Foundation Hospital, Urbana, Illinois. The tumors were diagnosed as invasive ductal carcinomas via standard pathological evaluation techniques performed at Carle Foundation Hospital. After surgical excision, one tissue sample from each patient was placed in a buffered saline solution and stored in a refrigerator or cooler until imaging was performed, typically less than $6\mathrm{h}$ later. OCT data were acquired from the margin region, the interface between normal and diseased tissues, and from tumor, adipose, and stroma tissues individually, having been identified and separated in the laboratory based on visual and tactile inspection. The assessment of tissue properties within each image was performed in the laboratory. To facilitate registration between OCT images and histology, the scanned tissue region was marked with India ink, fixed in 10% formalin solution, embedded, sectioned, and stained with hematoxylin and eosin (H&E) for evaluation using a light microscope.

The acquired OCT scan lines have a pixel size of $2\mu \mathrm{m}$ and extend to approximately $1.3\text{to}1.5\mathrm{mm}$ in depth. The preliminary data presented here are composed of 1666 scan lines of adipose tissue, 1408 scan lines of tumor tissue, and 941 scan lines of stroma tissue. The data were acquired using an OCT system employing a neodymium:vanadate pumped titanium:sapphire laser with a center wavelength of $800\mathrm{nm}$ and a bandwidth of $70\mathrm{nm}$ , yielding an approximate axial resolution of $4\mu \mathrm{m}$ . Light was introduced into a single-mode $50∕50$ fiber optic splitter (Gould Fiber Optics, Inc.), which coupled light from the source to sample and reference arms, and then back onto a silicon photodetector (New Focus, Inc., Model 2007). Approximately $10\mathrm{mW}$ of sample arm light was focused into the tissue by a $20\text{-}\mathrm{mm}$ achromatic lens, yielding a $15\text{-}\mu \mathrm{m}$ spot size (transverse resolution). Spatial scanning over the sample was achieved using a galvanometer-mounted mirror (Cambridge Technology, Inc). The reference arm comprised a galvanometer-based delay operating at 30 lines per second. Sampling and data acquisition were performed with a dedicated computer card (National Instruments, Models PCI-6110, PCI-6711) having a $10\text{-}\mathrm{MHz}$ sampling rate and a $12\text{-}\text{bit}$ quantizer. The logarithm of the amplitudes of the acquired data were displayed and stored on a personal computer.

The OCT response from adipose tissue was characterized by spatially periodic highly scattering boundaries due to the large size of the lipid-filled adipocytes $\left(60\text{to}120\mu \mathrm{m}\right)$ ,³⁶ while stroma tissue was more densely scattering and tumor tissue even more so. These properties are clearly evident in the representative adipose, stroma, and tumor axial scans $r_{\mathit{\text{tissue}}}\left(z\right)$ shown in Fig. 1 . The averaged Fourier-domain responses $R_{\mathit{\text{tissue}}}\left(k\right)=FFT\left[r_{\mathit{\text{tissue}}}\left(z\right)\right]$ from the adipose, tumor, and stroma tissues are shown in Fig. 2 . Attenuation effects have been removed from these data by subtracting a first-order linear fit from each response and every scan line has been normalized by the total integrated response over all spatial frequencies $k$ . Additionally, each scan line has been truncated so that only the region below the tissue surface remains, thereby removing effects arising from the varying free-space distance above the sample surface. Each tissue type displays a unique Fourier-domain signature that may be used to differentiate between scan lines of different tissue types.

Fig. 1

OCT axial scan data from human breast tissue (logarithmic response). (a) The periodic response due to adipose tissue. (b) The dense scattering response and associated attenuation effects from invasive ductal carcinoma tumor tissue. (c) The dense scattering response from stroma tissue with less prevalent attenuation effects and higher frequency oscillations than in tumor tissue.

Fig. 2

Fourier-domain data from OCT axial scans showing the unique signature from each tissue type. The averaged responses are shown over all tumor (black), stroma (dark gray), and adipose (light gray) tissue axial scans normalized by total area.

Based on the spatial data, it was also expected that the mean distance between high-intensity reflections in each scan would vary between tissue types. Therefore, we measured the mean distance between regions of intensity greater than half of the maximum in each scan line. Figure 3 shows the histogram of the distance measurements from all scan lines. These data show that adipose tissue exhibits a larger mean distance due to the large adipocyte cell boundaries while stroma and tumor tissues exhibit a smaller mean distance due to the smaller cell sizes and denser clustering of scattering objects. These data yield a simple means of classification via comparison of an unknown periodicity response with a set of known values. It is also clear from these data, however, that this property will not allow for highly effective differentiation between tumor and stroma tissues.

Fig. 3

Histogram of the mean distance between high-intensity backreflections in OCT data from human tumor (black), stroma (dark gray), and adipose (light gray) tissues. Each datum in the histogram is derived from one scan line.

2.2.

Computational Techniques

Two techniques were evaluated for the identification of tumor tissue. The first takes advantage of the unique Fourier-domain signature from each scan line, and the second exploits the periodic scattering response from adipose tissue. Each scan line, or image column, was classified individually in order to facilitate the use of these techniques in conjunction with the various forward-imaging devices present in the literature, which may yield single forward-directed scan lines of data, create circular images by combining scan lines from a radially directed and angularly scanned beam, or use various beam translation schemes to form traditional OCT B-scan images.^{22, 37, 38} The techniques presented here may also be applied to layered tissue boundaries by evaluating each scan line over a windowed region, which may be shifted along the axial response.

The Fourier-domain classification process was implemented by comparing the Fourier-domain data from each unknown scan line $F\left(k\right)$ with the averaged adipose, tumor, and stroma responses $R_{\mathit{\text{tissue}}}\left(k\right)$ from a set of training data. The tissue comparison with the lowest cumulative error

The periodicity analysis technique was implemented by comparing the mean distance between high-intensity reflections in a scan line of unknown tissue $d$ with a set of known values. The known tissue having the maximum histogram intensity $H_{\mathit{\text{tissue}}}\left(d\right)$ at this value yields the most probable classification. In this analysis, high-intensity regions of each scan line were constrained to those greater than half of the maximum scan line intensity and greater than $4\mu \mathrm{m}$ in width, reducing noise effects. Figure 4 shows the processing steps for both techniques.

Fig. 4

Diagram of the processing steps implemented in the Fourier-domain and periodicity analysis techniques.

A combination of these two techniques was also implemented. The classification of each scan line was assigned a confidence rating for both evaluation techniques. The Fourier-domain and periodicity techniques generated confidence ratings ( $c_F$ and $c_P$ ) by calculating the difference between the cumulative error or histogram amplitude, respectively, of the classified tissue type and the next-best tissue type.

Studies of layered tissue, those having multiple tissue types in a single scan line, may also be accomplished using these techniques. This analysis relies on the use of a windowing scheme, wherein sections of the axial data are evaluated separately. For example, a $100\text{-}\mu \mathrm{m}$ window may be scanned along a $1\text{-}\mathrm{mm}$ depth scan of data in $10\text{-}\mu \mathrm{m}$ steps. After evaluation of the data in each window region via the Fourier-domain, periodicity, or combined technique, overlapping results in each $10\text{-}\mu \mathrm{m}$ region are compared, with the most common classification being selected. These results will be limited by the relationship between window size and feature size. For instance, if the window size is too small to capture the entire response from an adipocyte, the periodicity technique will not produce a measurement matching the expected response. However, too large a window size will result in a response that is not fine enough to accurately detect boundaries between tissue layers. Fourier-domain results are especially influenced by time-frequency tradeoffs, most notably in the low-frequency region that is critical to this classification technique.^{39, 40} To combat these limitations, future work may be undertaken to analyze the performance of more complex techniques employing variable window lengths.