Diagnostic potential of multimodal imaging of ovarian tissue using optical coherence tomography and second-harmonic generation microscopy

Abstract.

Ovarian cancer is particularly deadly because it is usually diagnosed after it has metastasized. We have previously identified features of ovarian cancer using optical coherence tomography (OCT) and second-harmonic generation (SHG) microscopy (targeting collagen). OCT provides an image of the ovarian microstructure, while SHG provides a high-resolution map of collagen fiber bundle arrangement. Here, we investigated the diagnostic potential of dual-modality OCT and SHG imaging. We conducted a fully crossed, multireader, multicase study using seven human observers. Each observer classified 44 ex vivo mouse ovaries (16 normal and 28 abnormal) as normal or abnormal from OCT, SHG, and simultaneously viewed, coregistered OCT and SHG images and provided a confidence rating on a six-point scale. We determined the average receiver operating characteristic (ROC) curves, area under the ROC curves (AUC), and other quantitative figures of merit. The results show that OCT has diagnostic potential with an average AUC of $0.91\pm 0.06$. The average AUC for SHG was less promising at $0.71\pm 0.13$. The average AUC for simultaneous OCT and SHG was not significantly different from OCT alone, possibly due to the limited SHG field of view. The high performance of OCT and coregistered OCT and SHG warrants further investigation.

1. Introduction

Ovarian cancer is the deadliest cancer of the female reproductive system. While the prevalence of ovarian cancer is low (the lifetime risk in the United States is $\sim 1.4\%$), the majority of patients with the disease will die from it. This is because 61% of diagnoses are made after the disease has metastasized, and the five-year survival rate after metastasis is just 27%. By comparison, the survival rates at the localized and regional stages are 92 and 72%, respectively.¹ Biopsy during laparoscopy is the gold standard for diagnosis, but the procedure is too invasive to be a routine screening procedure. An effective screening procedure capable of detecting ovarian cancer at an early stage might greatly reduce the mortality rate of this disease.

Unfortunately, two of the most common screening methods for ovarian cancer, CA-125 blood test and transvaginal sonography (TVS), have not reduced ovarian cancer mortality compared to routine medical care.² Research has shown elevated serum levels of CA-125 in women with ovarian cancer;³ however, other factors also affect CA-125 levels, leading to false positives.⁴ TVS can detect some morphological changes of the ovaries associated with ovarian cancer. The procedure is fast and safe, which enables it to be a common screening procedure. Unfortunately, TVS suffers from moderate sensitivity and a low positive predictive value (PPV) (85.0 and 14.01%, respectively, from a study of annual TVS screening on 25,327 women).⁵ TVS performs especially poorly when the ovaries undergo only minor volume change due to cancer.⁵ The inability of TVS to resolve cellular or microstructural features in ovarian tissue is an obstacle that may prevent common diagnosis of early-stage ovarian cancer. Optical imaging techniques may be attractive alternatives to TVS due to their superior resolution.

Optical imaging has been shown to resolve microstructural or cellular features in ovarian tissue and to obtain biochemical information through the use of specific or nonspecific fluorescent contrast agents, or through interrogating endogenous fluorophores. These capabilities may be used to guide biopsy, the current gold standard for cancer diagnosis, by highlighting suspicious regions. The obvious disadvantage of optical techniques is the requirement of close proximity of the imaging probe to the tissue, thereby making these procedures more invasive than TVS. However, the potential for superior diagnostic accuracy makes optical techniques worthy of investigation as screening procedures. Upon identification of highly effective optical imaging techniques for ovarian cancer diagnosis, our attention can be turned to the necessary engineering to reduce the invasiveness of the imaging system to enable widespread and routine use in the clinic.

Several different optical modalities have been used to investigate ovarian tissue, including fluorescence spectroscopy,⁶ multispectral imaging,⁷ confocal microlaparoscopy,⁸ and optical resolution photoacoustic microscopy.⁹ Two optical techniques that have shown promise in the identification of abnormal ovarian tissue, but do not require exogenous fluorophores, are optical coherence tomography (OCT) and second-harmonic generation (SHG) microscopy. OCT generates cross-sectional images similar to TVS, but with micron-scale resolution.¹⁰ Near-infrared light penetrates the tissue and some of the light reflects at interfaces of refractive index. By scanning over a region of tissue, OCT generates two- or three-dimensional images of the anatomical structure of the tissue to a depth up to 2 mm. Current commercial OCT systems have lateral and axial resolutions better than 10 μm. This high-resolution capability allows OCT to detect minute changes in cellular layer thickness or other microstructural changes. Three-dimensional volumetric images can be generated in a few seconds.

SHG is a form of multiphoton microscopy in which contrast is formed by the nonlinear scattering of incident light by molecules with noncentrosymmetric structure.¹¹ The nonlinear nature of the scattering in SHG results in emission light with double the frequency of the excitation light. Scanning the focused excitation beam across the tissue generates three-dimensional volumetric images. Optical sectioning occurs because the emission intensity falls off with distance from the beam focus to the fourth power. The necessity of high numerical aperture to generate efficient second-harmonic emission yields lateral resolutions $<1\mu \text{m}$.

Previous research has shown that OCT and SHG can identify microstructural changes in ovarian tissue due to ovarian cancer development.^12,13 We have identified structural features characteristic of ovarian cancer using OCT in mice,¹⁴ rats,¹⁵ and humans.¹⁶ In healthy tissue, follicles of various stages of development and corpora lutea are clearly visible. The texture of the surrounding tissue appears mostly homogeneous and signal attenuation occurs slower than in abnormal tissue. Furthermore, OCT can visualize inclusion cysts and epithelial invaginations, which are thought to be risk factors for ovarian carcinoma.¹⁷ In SHG images of healthy tissue, the collagen structure appears highly ordered and the fibrils are thin. Conversely, in cancer tissue, the fibrils appear thick, wavy, and bunched together with an overall disordered structure.¹⁸ We have previously analyzed SHG collagen structure in mouse ovarian tissue using image processing algorithms to generate parameters for comparison and classification of normal and neoplastic ovarian tissue.¹⁹ Because SHG and OCT detect different features associated with ovarian cancer, combining the two modalities provides complementary information that might increase diagnostic performance compared to either modality alone.

The purpose of this study was to determine whether multimodal imaging consisting of OCT and SHG outperforms either modality alone for the purpose of detecting abnormalities in a mouse model of ovarian cancer. Naïve human observers classified each tissue volumetric image as normal or abnormal. We assessed the performance of the observers and the diagnostic accuracy of OCT and SHG for this task using receiver operating characteristic (ROC) analysis²⁰ and other quantitative figures of merit.

2. Materials and Methods

2.1. Images

The image data set consisted of previously collected SHG and OCT volumetric images of 44 ex vivo mouse ovaries.¹⁹ In the previous study, the mice were dosed with 4-vinylcyclohexene diepoxide (VCD) to simulate menopause through follicular apoptosis and then with the carcinogen 7,12-dimethylbenz[a]anthracene (DMBA) to induce cancer. Four experimental groups were used: VCD only, DMBA only, VCD and DMBA, and neither VCD nor DMBA. A vehicle (sesame oil) only was in the control groups. The ovaries were harvested five or seven months after DMBA injection and imaged using OCT and SHG. The OCT images were collected using a swept-source system with a center wavelength of 1040 nm and a bandwidth of 80 nm (OCS1050SS, Thorlabs, Newton, New Jersey). The resolution was 12 μm laterally and 9 μm axially in tissue. The SHG images were collected with a multiphoton microscope (TrimScope, LaVisionBioTec, Bielefeld, Germany) with a titanium-sapphire laser (Chameleon Ultra2, Coherent, Santa Clara, California). The excitation wavelength was 780 nm and the lateral resolution was 0.5 μm. The OCT field of view was $3\times 3\times 2\mathrm{mm}$ ($\text{lateral}\times \text{lateral}\times \text{depth}$), which was large enough to completely encompass most of the ovary samples. This is one benefit of imaging mouse ovaries; human ovaries are $\sim 4\mathrm{cm}$ long by 2 cm wide,²¹ which is too large to fit in the field of view of our OCT probe. The SHG field of view was $400\times 400\mu \text{m}$ ($\text{lateral}\times \text{lateral}$), with discrete depth information obtained from slices at 10-μm increments to depths up to 100 μm. The SHG field of view was coregistered to the center of the OCT field of view, which usually, but not always, corresponded to the center of the tissue sample. The volumetric data could be displayed as 700 cross-sectional or 512 en face (i.e., the plane parallel to the tissue surface) image slices for OCT, or 10 to 20 en face slices for SHG. The observers could view the OCT data in the two cross-sectional planes or the en face plane. They could only view the SHG data en face.

To determine accurate information about the diagnostic value of imaging modalities for specific tasks, it is important that the chosen cases are not too easy or too difficult to classify.²² We included all the images from the study in Ref. 19 except for those volumetric images that had very low image quality. For OCT, we omitted images where the tissue border was not clearly distinguished from surrounding noise, which corresponded to a signal-to-noise ratio of $<7$. In order to keep all OCT and SHG images paired for the combined OCT+SHG trials, we omitted SHG volumetric images whenever their corresponding OCT volumetric images were removed for poor quality.

2.2. Histopathological Analysis

Histopathological analysis was used to yield the truth values in this study.¹⁹ The tissue samples were embedded in paraffin after imaging and sectioned in 5-μm-thick slices perpendicular to the tissue surface (cross-sectional sections). Every 20th section was stained with hematoxylin and eosin. Of the 44 tissue samples, 16 were classified as normal and 28 as abnormal by both a pathologist and a gynecologic oncologist with veterinary training. The diagnosis distribution is shown in Table 1. VCD/DMBA effect was defined as morphological changes due to VCD and/or DMBA dosing. Features of VCD/DMBA effect were proliferation of the epithelium and degeneration of follicles and corpora lutea. Tubular adenoma is a benign tumor with the potential to become cancerous. Carcinoma includes granulosa cell tumor, Sertoli-Leydig cell tumor, and adenocarcinoma.

Table 1.

Distribution of truth values.

	Abnormal
Normal	4-vinylcyclohexene diepoxide/7,12-dimethylbenz[a]anthracene effect	Tubular adenoma	Carcinoma
16	6	19	3

2.3. Participants

Seven naïve observers were recruited from undergraduate and graduate students at the University of Arizona. By self-report, each observer had natural or corrected 20/20 vision and no experience reading OCT or SHG images.

2.4. Observer Trial Procedures

Each observer participated in three sessions corresponding to the three imaging modalities (OCT, SHG, and simultaneous OCT+SHG) separated by at least one week to reduce the likelihood that they would remember their perceived relative proportion of normal to abnormal classifications.²² Every observer rated every volumetric image in all modalities, resulting in a fully crossed, multireader, multicase (MRMC) study design.

Prior to each session, each participant underwent a two-part training procedure. In the first part, each observer passively viewed a training video recorded by the investigators. The videos contained examples from the relevant modality of normal and abnormal ovaries and described the important features in each volumetric image. By using videos, the overall training experience received by each observer was similar. The video for the OCT session also briefly described ovarian anatomy. Each video showed three example volumetric images (one normal and two abnormal). Figures 1 and 2 depict slices from each of the example tissue samples in the training videos. The video for the final session employing combined OCT+SHG incorporated the same examples from the previous two sessions. Following the videos, each observer had the opportunity to ask the investigator any questions about the video and task.

Fig. 1.

Slices from the optical coherence tomography (OCT) volumetric images shown in the training videos. The top row consists of en face images, the middle row consists of cross-sectional images, and the bottom row consists of cross-sectional images orthogonal to the images in the middle row. All three images in each column are from the same tissue sample. Normal ovaries [(a), (d), and (g)] in OCT are characterized by the presence of follicles and corpora lutea, relatively homogeneous texture in (a), and slow signal attenuation with depth and smooth tissue surface in (d) and (g). The characteristics of abnormality in (b), (e), and (h) are the inhomogeneities of the texture in (b), the rapid signal attenuation with depth in (e) and (h), the multiple small, dark cavities in (e) and (h), the rough tissue surface in (e) and (h), and the lack of follicles and corpora lutea. In the second abnormal example [(c), (f), and (i)], the abnormal characteristics are the inhomogeneous texture in the lower half of (c), the inhomogeneous rate of attenuation in (f) and (i), and the absence of follicles and corpora lutea. The bar in (a) is 1 mm and applies to each image.

Fig. 2.

Slices from the second-harmonic generation (SHG) volumetric images shown in the training videos. The thin, straight, organized collagen strands in (a) are characteristic of normal ovarian tissue in SHG. The abnormal characteristics in (b) are the hazy signal and the small, round spots. The signal in the left half of (c) is thick and unorganized, which is characteristic of abnormal ovarian tissue in SHG. The bar in (a) is 100 μm and applies to each image.

The second part of the training consisted of an active exercise in which each observer used the software to view and classify three or six volumetric images from that session’s image set as normal or abnormal. During this exercise, each observer was asked to explain their decision, and they were asked to point to any suspicious regions if they believed the tissue was abnormal. The investigator provided immediate feedback by confirming the features that were correctly identified by the observer, as well as explaining why any incorrect interpretations were wrong. The training images were the same for all observers. If any observer felt that they needed additional practice beyond the initial three volumetric images, this exercise was repeated on three additional volumetric images. The three (or six, if necessary) volumetric images from this exercise were included in the experimental data set. After completing this exercise, each observer began the trial and was not permitted to interact with the investigator.

The software used by the observers was written in MATLAB® (R2012a, MathWorks, Natick, Massachusetts). A screenshot of the software used on the third session is shown in Fig. 3. The OCT data are on the left and the SHG data are on the right. The buttons on the lower left control the camera for the OCT data. While the OCT data are shown in an isometric view in this figure for visual effect, the observers examined the images with the view oriented normal to one of the three orthogonal planes. The three sliders cycle through the OCT slices in the corresponding planes. The SHG data could only be viewed from the en face perspective, so only one slider, located below the image, is used to navigate the volume by viewing slices at different depths. After viewing each volumetric image or volumetric image pair, the observers made a binary classification decision of Normal or Abnormal and denoted their level of confidence on a three-point ordinal scale (Possibly, Probably, or Definitely). Once the observers confirmed their decision, the software displayed the next volumetric image in the set. Observers were not permitted to return to previously classified images. All observers viewed the same volumetric images in the same order. A random-number generator was used to randomize this order. The observers were instructed to not assume any particular prevalence of abnormal ovaries.

Fig. 3.

Screenshot of the software used by the observers in the third session (OCT+SHG). An example OCT volumetric image is shown on the left and its corresponding SHG volumetric image is on the right.

2.5. Analysis

Several quantitative figures of merit to describe the image quality of OCT, SHG, and combined OCT+SHG for the classification task were calculated. Using the binary response data, we calculated sensitivity, specificity, PPV, and negative predictive value (NPV) for each modality, averaged over all observers. These are defined below, where TP, TN, FP, and FN are true positive, true negative, false positive, and false negative, respectively.

$$\text{sensitivity}=\frac{\mathrm{TP}}{\mathrm{TP}+\mathrm{FN}},$$ (1)

$$\text{specificity}=\frac{\mathrm{TN}}{\mathrm{TN}+\mathrm{FP}},$$ (2)

$$\mathrm{PPV}=\frac{\mathrm{TP}}{\mathrm{TP}+\mathrm{FP}},$$ (3)

$$\mathrm{NPV}=\frac{\mathrm{TN}}{\mathrm{TN}+\mathrm{FN}}\mathrm{.}$$ (4)

The standard 95% confidence interval formula for binomial random variables, $\hat{p}\pm 1.96n^{-1/2}{[\hat{p}(1-\hat{p})]}^{1/2}$, is not valid for $\hat{p}=0$ or 1, where $\widehat{p}$ are the proportions of true positives to total abnormal population for sensitivity, true negatives to total normal population for specificity, true positives to number of abnormal classifications for PPV, and true negatives to number of normal classifications for NPV. There were several instances in this study where $\hat{p}=1$, therefore, the 95% confidence intervals were determined using the Wilson interval, which is more accurate and more valid.²³

$$\frac{1}{1+\frac{z^2}{n}}[\widehat{p}+\frac{z^2}{2n}\pm z\sqrt{\frac{1}{n}\widehat{p}(1-\widehat{p})+\frac{z^2}{4n^2}}],$$ (5)

where $z$ is the z-score (1.96 for 95% confidence intervals) and $n$ is the sample size. Averages of these figures of merit and their average confidence intervals were calculated using the arithmetic mean.

Using the observers’ confidence data, we used ROC analysis to determine the diagnostic performance of the observers and of the imaging modalities. ROC analysis is based on statistical decision theory and provides the most comprehensive description of the sensitivity and specificity of a diagnostic system. Furthermore, the area under the ROC curve (AUC or $A_z$) provides a single quantitative measure of the diagnostic performance of a system for a given task.^20,22 While it is possible for different ROC curves to have the same AUC, the AUC is an unambiguous metric for comparison in cases where the compared ROC curves do not cross. The confidence data used in the ROC analysis were formed by mapping the binary and three-point ordinal confidence data to a six-point scale according to Table 2.

Table 2.

Score values for receiver operating characteristic (ROC) analysis.

Scores	Interpretation
0	Definitely normal
1	Probably normal
2	Possibly normal
3	Possibly abnormal
4	Probably abnormal
5	Definitely abnormal

For each modality, we plotted the fitted average ROC curve and calculated the average AUC. The fitted average ROC curves were plotted using the means of the $a$ and $b$ parameters of the binormal distribution from each observer-modality pair,^22,24 where the parameters are defined as

$$a=\frac{{\widehat{\mu }}^{+}-{\widehat{\mu }}^{-}}{{\widehat{\sigma }}^{+}}b=\frac{{\widehat{\sigma }}^{-}}{{\widehat{\sigma }}^{+}},$$ (6)

where $\widehat{\mu }$ and $\hat{\sigma }$ are the estimated distribution mean and standard deviation, respectively, and the positive and negative superscripts denote the corresponding mode of the bimodal binormal distribution. The assumption that the distribution of scores, or a monotonic transformation thereof, fits a binormal distribution has been theoretically and empirically validated.²⁵ There are multiple methods to determine the $a$ and $b$ coefficients. The most accurate methods involve maximum-likelihood estimation to iteratively fit a binormal ROC curve to the data. However, sometimes these techniques fail to converge due to the form of the empirical data.²² Because some of our reader-modality ROC curves failed to be fit using maximum-likelihood methods, we instead directly determined the binormal parameters using the sample mean as an unbiased estimate of the population mean and standard deviation as a biased estimate of the population standard deviation (though some bias is reduced due to Bessel’s correction).²⁴ It should be noted that while this method of fitting the average ROC curves is parametric, the empirical observer-confidence ratings are nonparametric. As such, the smooth average ROC curves are provided for visualization purposes. Any quantitative comparisons of the imaging modalities should be made using the average AUCs.²⁴

Finally, the average AUC, AUC 95% confidence interval, and hypothesis testing were calculated using the iMRMC software (version 2.0b) for each modality.²⁶ The first null hypothesis was that the average AUC for each modality was equal to 0.5, where 0.5 corresponds to a system that guesses randomly. The other null hypotheses were that the average AUCs of any two modalities were equal. The $p$ values and confidence intervals were calculated using the modified Dorfman-Berbaum-Metz method for ROC analysis summarized by Hillis, Berbaum, and Metz.²⁷ The iMRMC calculations are made with U-statistics, so the AUCs will not exactly match the areas of the smooth, binormal curves.

3. Results

The average sensitivity, specificity, PPV, NPV, and AUC (plus or minus one standard deviation) for OCT, SHG, and simultaneous OCT+SHG are shown in Table 3. Sensitivity and specificity were determined using the binary classification information. This is equivalent to a threshold of 2.5 using the score definitions in Table 2. The mean values of each of the figures of merit were highest for OCT, followed by OCT+SHG. The SHG figures of merit were the lowest.

Table 3.

Figures of merit for each modality (95% CI lower bound, 95% CI upper bound).

Figure of merit	Optical coherence tomography (OCT)	Second-harmonic generation (SHG)	OCT+SHG
Area under the ROC curve	0.91 (0.85, 0.97)	0.71 (0.58, 0.84)	0.89 (0.80, 0.98)
Sensitivity	0.94 (0.80, 0.98)	0.78 (0.60, 0.89)	0.93(0.76, 0.96)
Specificity	0.88 (0.66, 0.96)	0.65 (0.42, 0.83)	0.72 (0.45, 0.84)
Positive predictive value	0.94 (0.79, 0.98)	0.80 (0.62, 0.91)	0.86 (0.67, 0.92)
Negative predictive value	0.91 (0.68, 0.98)	0.64 (0.40, 0.82)	0.88 (0.58, 0.92)

The fitted average ROC curves are shown in Fig. 4 and the AUC comparisons are shown in Fig. 5. All modalities performed significantly better than chance ($p<0.01$ for each). The average OCT AUC was greater than the average SHG AUC ($\mathrm{\Delta }\mathrm{AUC}=0.20\pm 0.15$ (95% CI), $p=0.011$). The average AUC for OCT+SHG was also greater than the average SHG AUC ($\mathrm{\Delta }\mathrm{AUC}=0.17\pm 0.13$, $p=0.011$). However, the average AUCs for OCT+SHG and OCT alone were not significantly different ($\mathrm{\Delta }\mathrm{AUC}=0.03\pm 0.08$, $p=0.496$).

Fig. 4.

Fitted average receiver operating characteristic (ROC) curves for each modality. Binormal ROC curves were fit to each observer trial to yield the binormal distribution parameters $a$ and $b$. These parameters were averaged over all observers to form an average binormal ROC curve for each modality.

Fig. 5.

Comparison of average area under the ROC curves (AUCs) for each modality. All average AUCs are significantly greater than chance ($p<0.01$). The average AUC for OCT was significantly greater than the average AUC for SHG ($p=0.011$). Likewise, the average AUC for OCT+SHG was greater than SHG ($p=0.013$) The average AUC for OCT was not found to be significantly different from OCT+SHG.

4. Discussion

The performance of OCT in classifying abnormal ovaries is promising. The average AUC shows that a nonexpert reader can be expected to correctly identify an abnormal ovary $>90\%$ of the time when averaged over all specificities.²⁸ Also, the PPV and NPV are both $>0.9$ for OCT, meaning that 90% of the time, a decision made by an observer given the particular distribution of positive and negative cases in this study will be correct. It should be noted, however, that the distribution in this study does not match the actual distribution of ovarian disorders that naturally occur in women. 86,138 women in the United States are currently estimated to have ovarian cancer ($\sim 0.12\%$ of the female population).¹ Tubular adenoma is extremely rare in women and the effects of DMBA dosing do not occur naturally. Therefore, the PPV and NPV reported here cannot be generalized to the clinical population.²⁴ A benefit of OCT is that several ovarian features can be analyzed—such as surface smoothness, attenuation, and texture variability—to make a classification decision. The large field of view is also helpful because the observers could view the entire ovary.

Observers reading SHG images correctly classified the tissue better than chance, but the overall performance was poorer than OCT. With SHG alone, readers can be expected to correctly classify abnormal ovaries greater than two-thirds of the time. The sensitivity was greater than the specificity, which suggests that the observers found signs of abnormality to be more noticeable than signs of normality. The overall diagnostic performance of SHG in this study was not shown to be a promising screening technique on its own, though there were some particular limitations to the SHG data we used.

There are a few possible explanations for the poor performance of SHG. The most likely shortcoming of the SHG data in this study is the limited field of view. The SHG images were taken over a $400\times 400\mu \text{m}$ region centered on the center of the $3\times 3\mu \text{m}$ OCT field of view. By limiting the field of view, the abnormal collagen structures of VCD/DMBA effect and carcinoma samples might be difficult to detect, since these structural changes could be most easily noticed when observing the collagen structure over the entire ovary. We envision that with a clinical dual-modality endoscopic OCT-SHG system, the clinician would obtain SHG images at several regions in an ovary guided by the gross OCT data. The smaller amount of SHG data compared to OCT may also limit the performance of SHG in this study. Each OCT volume contained two orders of magnitude more slices in the depth direction than each SHG volume. This difference in data set size was mainly attributable to speed. The OCT system was able to obtain near-isotropic pixel-size image sets usually encompassing the entire lateral extent of the ovary in 30 s, whereas the SHG system required 15 s per image. Therefore, a limited number of slices were obtained at a relatively large slice separation (10 μm) much greater than the image axial resolution (1.2 μm). This shortcoming could be improved with the use of higher incident light power and faster scanning. The low incident power (20 mW) was chosen as a safe power level to mimic the necessarily low power levels required for in vivo imaging. SHG also had lower depth of penetration, with signal loss at $\sim 100\mu \text{m}$ depth, whereas OCT was able to image $\sim 1\mathrm{mm}$ deep. We and others have created metrics for SHG imaging of collagen with statistically significant differences between healthy and malignant ovarian tissue. Some examples include Fourier analysis,^19,29 gray-level co-occurrence matrices,¹⁹ SHG emission intensity and directionality,³⁰ and the distribution of collagen fiber angles relative to the epithelial boundary.²⁹ Therefore, we believe that addressing the SHG shortcomings in this study may improve the observer diagnostic performance of a dual-modality OCT-SHG imaging system.

When comparing the imaging modalities, the average AUCs show that OCT is superior to SHG for classifying ovarian tissue as normal or abnormal using nonexpert observers. Interestingly, we also found that observers did not perform significantly differently from OCT when they could simultaneously view SHG images of the same tissue. The average AUC for the multimodal trial should have been higher than OCT alone if the observers came to their decisions primarily based on the OCT data, but were able to strengthen their confidence with the SHG data. This suggests that the extra information provided by the SHG images was either not always used or was possibly misleading. It is possible that some of the dual-modality images may appear normal in OCT and abnormal in SHG or vice versa potentially due to the limitations of the SHG images. Again, performance of the multimodality trial might be improved in future studies by collecting high-density SHG images across the entire OCT field of view.

A successful candidate for a new ovarian cancer screening procedure must exhibit high sensitivity and specificity for any reader. The observers’ binary classification data for OCT resulted in a sensitivity of 0.94 and specificity of 0.88. A controlled trial in the United Kingdom of $>50,000$ women found transvaginal sonography to have a sensitivity of 84.9% and a specificity of 98.2% for the detection of ovarian and tubal cancers.³¹ An American study of $>25,000$ women screened annually with TVS yielded a sensitivity of 85.0% and a specificity of 98.7% (histopathological analysis provided the truth values for both studies).⁵ These results suggest that TVS can discriminate ovarian cancer from healthy tissue reasonably well, but this does not mean they have high clinical value. The PPVs for TVS were 5.3 (Ref. 31) and 14.01% (Ref. 5) for the U.K. and U.S. trials, respectively. Since the prevalence of ovarian cancer is very low, TVS yields many more false positives than true positives (true positives would equal false positives for $\mathrm{PPV}=50\%$). Because sensitivity and specificity are independent of prevalence, they overestimate the clinical value of a test for diseases with low prevalence in the general population. Our discriminatory OCT results (sensitivity and specificity) are promising and comparable to the performance found in the U.K. and U.S. trials, but a study of the diagnostic performance of OCT and SHG on ovarian cancers with a disease distribution that matches the natural prevalence would provide the most useful information regarding the clinical value of such an imaging system. We could not directly compare OCT and SHG to TVS using a mouse model because the resolution of TVS ($\sim 0.25\mathrm{mm}$) is insufficient to resolve key anatomical features of the mouse ovary, such as oocytes or immature follicles.³² An important point of optical imaging techniques is that they can distinguish abnormalities, such as tubular adenoma, which may precede cancer. This means that survival might be improved. As such, OCT imaging for ovarian cancer screening warrants further investigation.

A future observer study should be conducted using in vivo images of human ovaries, preferably obtained by an endoscopic imaging system, in order to gain a more accurate measure of the diagnostic accuracy of OCT and SHG as they would be used clinically. Furthermore, future investigations of SHG should cover a larger area of the ovarian tissue. The performance of expert readers could also be evaluated; although given that both OCT and SHG are developmental technologies for this application, few expert readers currently exist. Finally, the distribution of abnormalities should be designed to closely match the distribution seen in the female population.

Biographies

Weston A. Welge is a doctoral student at the College of Optical Sciences at the University of Arizona. He received his BS in electrical and computer engineering and BA in history from the University of Colorado at Boulder and MS in optical sciences from the University of Arizona. His research focuses on the development of miniature optical endoscopes and novel applications of optical coherence tomography for early cancer detection.

Andrew T. DeMarco holds a bachelor’s degree in linguistics and a master’s degree in communication sciences and disorders from Temple University. He holds a certificate of clinical competence from the American Speech-Language Association. He is currently a doctoral candidate in the Department of Speech, Language, and Hearing Sciences at the University of Arizona, where he uses neuroimaging to understand the organization of language in the brain and how that architecture breaks down in aphasia after stroke.

Jennifer M. Watson has bachelor’s degrees in mechanical engineering and health sciences, and a PhD in biomedical engineering with a minor in entrepreneurship. She completed her dissertation research on evaluation of ovarian tumor development using the optical imaging modalities multiphoton microscopy and optical coherence tomography. In her current position at Tech Launch Arizona, she works with researchers across campus to assess the commercial potential of technologies, create development and commercialization plans and access key resources.

Photini S. Rice holds an associate of applied science degree in medical technology and is ASCP and HEW certified. She worked in the clinical laboratory setting for 10 years and consulted for 3 years at a physician’s office laboratory while working in a university research laboratory. She has spent the last 16 years working in cardiovascular and cancer imaging research at the University of Arizona.

Jennifer K. Barton is currently a professor of biomedical engineering, electrical and computer engineering, and optical sciences at the University of Arizona. She has served as the assistant director of the BIO5 Institute and interim vice president for research. She is a fellow of SPIE and the American Institute for Medical and Biological Engineering. Her work on development of miniature multimodality optical endoscopes and light-tissue interaction has led to over 90 peer-reviewed journal papers.

Matthew A. Kupinski is an associate professor in the College of Optical Sciences and the Department of Medical Imaging at the University of Arizona. He obtained his PhD from the University of Chicago in 2000. His work focuses on using objective, task-based measures of image quality to aid in the design of imaging systems.