2.1.

Monoculture and Coculture

CP-A (non-dysplastic metaplasia, ATCC® CRL-4027™) and CP-B (high-grade dysplasia, ATCC® CRL-4028™) are hTERT immortalized human Barrett’s esophageal cell lines.^22–24 In order to provide the ground truth for the classification results in coculture, parental CP-A cells were further transfected with a cytosolic mCerulean3 construct to achieve cyan fluorescence, which offers the advantage of not overlapping with the PpIX spectrum. The stably transfected CP-A cells were expanded using the complete culture medium containing $5\mu \mathrm{g}/\mathrm{ml}$ selective antibiotics (Blasticidin S HCl, R210-01, Gibco®, Canada). As some nonfluorescent cells may develop resistance to the selective antibiotics, a flow cytometer (FACSCanto™, BD Biosciences, California) was used to further purify mCerulean3 expressing cells. Stable cells were then maintained using complete medium supplemented with $3\mu \mathrm{g}/\mathrm{ml}$ of Blasticidin S.

The complete medium used for all cell lines was composed of MCDB 153 base medium (M7403, Sigma), supplemented with 5% fetal bovine serum (16000-036, Gibco®), $0.25\mu \mathrm{g}/\mathrm{ml}$ amphotericin B (A2942, Sigma), 1% penicillin/streptomycin (15140-122, Gibco®), $0.4\mu \mathrm{g}/\mathrm{ml}$ hydrocortisone (H0888, Sigma), $140\mu \mathrm{g}/\mathrm{ml}$ bovine pituitary extract (P1476, Sigma), $20\mathrm{mg}/\mathrm{l}$ adenine (A8626, Sigma), 0.1% insulin-transferrin-sodium selenite (I1884, Sigma), $20\mathrm{ng}/\mathrm{ml}$ recombinant EGF (E9644, Sigma), 4 mM glutamine, and 1 nM cholera toxin (C8052, Sigma). Cells were grown in $25{\mathrm{cm}}^2$ culture flasks and maintained at 37°C in a water jacketed ${\mathrm{CO}}_2$ incubator (Forma Series II, Thermo Fisher Scientific Inc., Waltham, Massachusetts). The subculturing procedures followed the recommendation from ATCC, where cells were immersed in 1 ml trypsin-EDTA solution (2.5X, 0.5%, 15400-054, Gibco®) until fully detached and neutralized by complete culture medium followed by centrifugation at 125 g in order to resuspend cells in fresh trypsin-free culture medium.

2.2.

Sample Preparation

A 100 mM stock solution of 5-ALA (A3785, Sigma) dissolved in phosphate buffered saline (14190-144, Gibco®) was stored in aliquots in the dark. A working concentration of 0.5 mM was prepared before every experiment by further diluting the stock solution in the serum-free complete culture medium. Cells were seeded on glass-bottom dishes (P35G-1.5-14-C, MatTek, Massachusetts) one day before the experiment. Before imaging acquisition, individual cell lines were incubated with 5-ALA for various periods of time between 3 and 6 h to yield different mitochondrial PpIX concentrations for training of the classification algorithm. This variable was intentionally introduced to characterize intracellular PpIX concentration and account for variability in fluorescence intensity during classification. Coculture preparation used a total of $0.3\times {10}^5\text{cells}$ containing mCerulean3-CP-A and parental CP-B in $1:1$ ratio with the incubation of 5-ALA for 6 h.

2.3.

Imaging Acquisition

The fluorescence emission of 5-ALA induced PpIX was collected between 625 and 750 nm by a laser scanning confocal microscope (TCS SP5 & DMI 6000 B, Leica, Wetzlar, Germany) equipped with an argon-ion laser operating at 514 nm. The emission spectral band starting from 625 nm was used to maximize the detected PpIX intensity while avoiding overlap with the mCerulean3 spectrum. A $63\times$ objective lens was used to yield a field of view of $246\mu \mathrm{m}\times 246\mu \mathrm{m}$ with lateral resolution of $\sim 300\mathrm{nm}$, comparable to current confocal endomicroscopy systems.¹² The acquisition channel for transfected CP-A cells in coculture used 458 nm laser excitation with the emission spectral range of 465 to 500 nm to match the fluorescence emission peak of mCerulean3. Sample fluorescence images of both cell lines in monoculture and coculture are demonstrated in Figs. 1 and 2.

Fig. 1

Sample CP-A and CP-B cells labeled with fluorescent protoporphyrin IX (PpIX) in monoculture conditions. Cells were loaded with 5-aminolevulinic acid (5-ALA) and the PpIX fluorescence was examined within 6 h of incubation. The fluorescent area (mitochondria distribution) is highlighted by the solid line, and the cell shapes based on the bright-field images are indicated by the dashed line. PpIX fluorescence yielded distinct morphological and textural patterns in (a) CP-A cells and (b) CP-B cells. CP-A cells yielded a more concentrated fluorescence accumulation toward the perinuclear region, while CP-B cells demonstrated a more elongated shape. Further image processing and segmentation was performed for classification of CP-B cells.

Fig. 2

Stably transfected CP-A cells were used for validation of the classification results. In the same region of co-culture shown in (a) and (b), (a) the CP-A cells expressing mCerulean3 are identified from the emission band of 465 to 500 nm, (b) thus serving as ground truth when both CP-A and CP-B (white arrows) are stained with 5-ALA induced PpIX.

2.4.

Quantifying Intracellular PpIX in Mono- and Coculture

To investigate the feasibility of using PpIX as a contrast agent, the cellular uptake was compared for metaplastic and HGD cells. Both cell lines were incubated with 5-ALA solution for a specified time and fluorescence images were acquired every hour using both one-photon ($5.92\mathrm{J}/{\mathrm{cm}}^2$) and two-photon excitation ($9911\mathrm{J}/{\mathrm{cm}}^2$) until the intracellular PpIX reached a plateau. PpIX intensities in coculture were compared only at 6 h of incubation. All acquisition parameters, including the fluence, magnification, emission spectral range, and the detector gain remained the same in repetitive trials using different batches of cells. The acquired images were preprocessed with consistent procedures to segment the fluorescent area using MATLAB® (MathWorks, Natick, Massachusetts). The average PpIX intensity per pixel was then plotted against the incubation period.

2.5.

Segmentation and Feature Extraction

The images were imported to MATLAB® for further image processing and feature extraction. Due to the low signal-to-noise ratio provided by PpIX, all fluorescence images were preprocessed to reduce the shot noise and uneven background. Contrast was enhanced by three iterative adjustments of imaging threshold, followed by further smoothing and morphological operations to produce segmented binary images of each cell. The segmentation results are demonstrated in Fig. 3. Then, mitochondrial distribution was analyzed using various morphological features based on the binary images to illustrate the shape eccentricity, irregularity, and the nucleus-to-mitochondria ratio.^9,27 Textural features, such as PpIX fluorescence patterns, of each cell were explored using calculation of the intensity histogram.²⁷ Further textural analysis was achieved using the gray-level co-occurrence matrix (GLCM)²⁸ along perpendicular and diagonal axes of the bounding box to extract intensity contrast, correlation, homogeneity, and energy. Note that the textural features were extracted from 8-bit intensity images. The key features and their descriptions are summarized in Table 2.

Fig. 3

Examples of morphological and textural features. (a) A binary image is used to extract morphological features. The bounding box, convex area (area within dashed lines), and the ellipse are noted. (b) The original 8-bit image of mitochondria distribution was obtained using the binary mask. This image shows the line used for extracting the line scan intensity profile as described in Table 2. (c) The intensity contrast can be represented by the slope of the cumulative intensity along the line. The slope (L) is from the line shown in Fig. 3(b), and the slope (R) is obtained from the line in the opposite direction.

Table 2

List of features

2.6.

Feature Reduction and Classification

An image dataset containing a total of 62 CP-B and 63 CP-A cells from monoculture were used to train the SVM. Morphological and textural features extracted from each individual cell are concatenated to form a 15-dimensional feature vector which characterizes the cells. Most of the features contain redundant information and are highly correlated. Therefore, reduction of feature dimensionality was first performed using a forward sequential selection (FSS) algorithm (sequentialfs, MATLAB®, MathWorks).²⁹ In the feature selection process, individual features were sequentially recruited and tested for their prediction accuracy using linear discriminant analysis with 10-fold cross-validation.³⁰ This operation splits the whole training set into 10 subgroups, followed by iterative classification on every single subgroup using the training algorithm obtained from the other nine-tenths of the datasets. This feature reduction step renders a combination of feature subsets that yields the smallest misclassification error [as shown in Fig. 5(a)].

Using the best feature subsets returned from the FSS algorithm, SVM were then trained to classify each cell in the coculture to CP-A (metaplasia) or CP-B (HGD). SVM mapped the input features to a higher-dimensional space to find the hyperplane that yields the largest margin between the two classes. The Gaussian radial basis function (RBF) kernel was applied to deal with potential nonlinear relationships between the feature attributes and the class, thus yielding a soft margin.³¹ Given the training sets, the RBF kernel radius ($R$) and the box constraint ($C$) were optimized using fivefold cross-validation. A total of 109 CP-A and 135 CP-B cells in coculture were then used for testing the classification performance. The performance of the SVM was measured by sensitivity and specificity. The sensitivity is defined as the ratio of correct HGD detection to true CP-B numbers, and the specificity corresponds to the ratio of correct metaplasia detection to true CP-A numbers. Performance of the classification algorithm was further assessed by plotting the receiver operating characteristic (ROC) curve and calculating the area under the curve (AUC), as shown in Figs. 5Fig. 6Fig. 7–8.

3.1.

Evaluation of Time-Lapse PpIX Fluorescence Intensity

To evaluate the use of 5-ALA for detection of dysplastic cells, we first characterized the cellular uptake of the two cell lines in monoculture. The intracellular PpIX intensities were plotted against the incubation time as shown in Figs. 4(a) and 4(b). The results showed that CP-B cells reached maximum PpIX concentration faster (within 3 h) than CP-A cells, where the PpIX intensities at 3 h were $32\pm 3$ counts in CP-B and $23\pm 2$ counts in CP-A using one-photon excitation ($p<0.01$). The data acquired using two-photon excitation demonstrated a similar trend. In addition, it was clearly shown that there was no selective accumulation of PpIX in HGD cells with neoplastic characteristics; instead, both cell lines exhibited similar saturation intensities after 5 h of 5-ALA incubation. As shown in Fig. 4(c), the lack of cellular uptake selectivity ($p=0.51$) was also observed in a coculture environment, where the PpIX intensities in both cell lines agreed with the monoculture results. Although we are comparing metaplasia versus HGD, the results are consistent with a previous study showing that BE tissue exhibited a faster PpIX accumulation within 4 h compared to normal epithelium, and the eventual maximum intensity was higher in normal epithelium.³² The conversion and saturation intensity of PpIX may depend on the activity and amount of ferrochelatase.^21,32

Fig. 4

(a) The average intracellular PpIX intensity of each cell line in monoculture was plotted against the incubation time from 0 to 7 h using one-photon (solid line) and two-photon excitation (dotted line). It was noted that CP-B cells exhibit a higher PpIX intensity than CP-A at 3 h. The intracellular PpIX intensity eventually reached a plateau. The standard deviation of each data point was not shown in the figure for easy visualization. (b) PpIX intensity of CP-B and CP-A in monoculture measured at 3 h ($p<0.01$) using one-photon excitation. (c) Intracellular PpIX intensities were compared between the two cell lines grown in coculture environment after 6 h of 5-ALA incubation ($p=0.51$). The error bars represent the standard deviation of the mean.

3.2.

Reduction of Dimensionality

The rank of features in the recruiting process using the FSS algorithm is indicated by the superscripts in Table 2. A lower value indicates a higher prevalence of the given feature. The examination of feature combination is then based on this sequence. During sequential feature recruiting performed with the monoculture training group, combinations of a range of feature numbers were examined for their classification performance. As shown in Fig. 5(a), the misclassification rates (MCR) were plotted against the number of features selected. The same combinations of features returned by feature selection were then tested in coculture classification. It is noted that the MCR converged to 0.10 when more than four features were used for the monoculture group, while only two features (eccentricity and line scan intensity profile) were required to produce the MCR to 0.09 when the same feature subsets were used in coculture classification. With the reduced feature dimensionality, the performance of SVM for coculture classification was further demonstrated in Fig. 5(b), where ROC curves were plotted for various feature numbers. The best prediction performance using two features achieved an AUC of 0.95. There is a trend that the error of coculture classification is almost doubled after three features. This could be due to the fact that the other features are not as discriminating and that they add noise to the process. Extraction of features can be a time-consuming task since it requires computing geometric, intensity, and texture features from each cell. Obtaining such a high AUC with only two features selected via sequential feature recruiting is a promising result since the SVM can use these two identified features alone for classifying CP-A versus CP-B cells in the context of 2-D coculture. From a computational point of view, this result opens the way for online implementation of CP-A–CP-B classification.

Fig. 5

(a) Misclassification rates were plotted against up to 10 features used for classification. The results from the training set using monoculture yielded a misclassification rate (MCR) of 0.10 when four features were used. It is noted that classification results using the same feature subsets in coculture yielded an MCR of 0.09 when only the eccentricity and line scanned intensity profile were employed. The error bars represent standard deviation of coculture testing sets from three separate acquisition trials. (b) Receiver operating characteristic curves were obtained from support vector machine (SVM) classification on all coculture samples, and an area under the curve (AUC) of 0.95 can be achieved using the two features returned by feature selection, as indicated by the bold black line.

3.3.

Classification Performance Using Support Vector Machines

Without any specified stopping criterion, the feature selection process returned two subsets: one comprised eccentricity and the slope of the line scan intensity profile, and the other comprised eccentricity and GLCM contrast. Figures 6 and 7 demonstrate the training results using monoculture cells and the corresponding classification results on the cocultured cells. Figure 6 shows the training results based on eccentricity and GLCM contrast using the SVM with a Gaussian kernel. The two optimized parameters for the kernel, $R$ (2.21) and $C$ (1.07), account for the radius of the Gaussian kernel and the imbalance of the group population, respectively. Although the training result demonstrates sensitivity of 90% and specificity of just 74%, the combined classification results from all coculture images of repetitive trials showed 94% sensitivity and 82% specificity. The ROC curves that demonstrate the trade-off of sensitivity and specificity are shown in Fig. 8. Figure 7 shows the training results based on eccentricity and the slope obtained from the line scan profile. This training result was obtained with a sensitivity of 95% and specificity of 70% with $R$ and $C$ of 2.17 and 0.78, respectively; however, the classification produced an improved prediction on the coculture set, for which 95% sensitivity and 87% specificity were obtained with an AUC of 0.95 (Fig. 8). Table 3 summarizes the classification results from both monoculture and coculture, indicating a good generalization capability of the classifier from monoculture to coculture environment.

Fig. 6

(a) The training results based on eccentricity and intensity contrast, which yield sensitivity of 90% (true CP-B) and specificity of 74% (true CP-A). (b) The trained model was tested on a combination of coculture imaging trials. The classification results demonstrated 94% sensitivity and 82% specificity.

Fig. 7

(a) The training results based on eccentricity and the slope obtained from the line scan profile using the same training set and testing set. This training showed sensitivity of 95% (true CP-B) and specificity of 70% (true CP-A). (b) The results from coculture images showed 95% sensitivity and 87% specificity.

Fig. 8

Performance of SVM using two features: the combination of eccentricity and intensity contrast ($\mathrm{AUC}=0.94$), and the combination of eccentricity and the slope obtained from the line scan intensity profiles ($\mathrm{AUC}=0.95$).

Table 3

Summary of the classification results.