2.1.

Bladder Anatomy

The urinary bladder consists of three layers: the epithelium (commonly called the urothelium) made up of transitional cells, a connective tissue layer known as the lamina propria, and the muscularis comprised of three concentric layers of muscle. In healthy tissue, these layers are well organized and well defined. The urothelium is the very thin innermost layer of the bladder wall, bound by the lamina propria with muscle the outermost tissue.

The primary growth patterns for urothelial neoplasms, which account for 98% of primary tumors of the bladder,⁶ are papillary lesions, CIS, and flat lesions. Papillary lesions are exophytic growths resembling a sea anemone, projecting into the bladder with multiple fronds. CIS is a diffuse, high-grade, intraepithelial malignancy. The much more uncommon flat lesions are heaped up epithelial tumors without the characteristic stalk of papillary lesions. Any of these tumors can become invasive by infiltrating the lamina propria and more deeply into the muscle.⁷ Of these lesions, only papillary lesions are readily visible endoscopically. In addition, with visual inspection it is not possible to determine if any of these lesions have become invasive.

Thickening of the urothelial layer by additional normal cells is known as hyperplasia and is usually associated with previous trauma or infection. Inflammation is classified as exudative or infiltrative, depending on the anatomic depth and type of cells involved in the inflammatory process. Cellular characteristics distinguish normal from malignant tissue with a spectrum of abnormal to precancerous conditions known as dysplasia between the two extremes. Dysplasia ranges from mild to severe with severe dysplasia often indistinguishable from CIS. While a thickened bladder epithelium may signify cancer, the cellular characteristics define its biological activity.⁸

2.2.

OCT Imaging of the Bladder

Feldchtein ⁹ demonstrated endoscopic use of OCT to examine the mucosal membranes of several internal organs including the bladder. They were able to image the discrete layers of the bladder, along with blood vessels and cysts located within these layers. In healthy tissue, the urothelium (U) appears in OCT images as an area of low intensity, while the lamina propria (LP) appears as an area of high intensity and the muscle (MS) appears as an area of low intensity thereby providing contrast between layers. Figure 1a shows an OCT image of the lining of a healthy bladder with well defined layers.

Fig. 1

OCT images of normal and cancerous bladder tissue, dimensions 1.8 by $1.5\mathrm{mm}$ , showing characteristics recognized by (a) Feldchtein ⁹ and (b) Jesser ¹⁰ Organized layers can be seen in the image of normal bladder tissue, while clearly defined layers cannot be seen in the image of the invasive tumor. The images were taken during a study at the George Washington University Medical Center¹¹ (GWUMC).

Jesser ¹⁰ used in vitro specimens to demonstrate the possibility of using OCT to distinguish between normal human bladders and those with invasive transitional cell carcinoma (TCC). Jesser noted that, unlike OCT images of the normal bladder, images of invasive TCC did not contain distinct layers or boundaries, and concluded that malignant invasion disturbed the normal well-defined strata seen in OCT images. Figure 1b shows an endoscopic OCT image of the bladder with an invasive tumor in which the layers cannot be distinguished.

In a study by Zagayanova, ¹² OCT scans from 63 patients were evaluated using OCT and compared with the pathological diagnosis obtained from biopsy. The OCT images were evaluated as being either malignant or benign, with the absence of a layered structure being the primary indicator of malignancy. The overall results had a sensitivity of 98% and a specificity of 72%. During a similar study conducted at our institution which also evaluated invasion,¹¹ 24 patients underwent cystoscopic examination of the bladder followed by OCT scan, photography, and biopsy from at least six mucosal areas. The OCT scans were reviewed and declared to be healthy, abnormal but not invasive, or invasive, and compared later with the pathological biopsy results. The results indicated 100% sensitivity and an overall specificity of 89%, with a negative predictive value of 100%. The algorithm introduced in this paper was tested on a subset of the images obtained during that study.

The preceding studies used human observers to analyze OCT images and assign diagnoses, which make the results highly dependent on the training and ability of the observers. A computer algorithm capable of analyzing and diagnosing the images would be operator–independent and make OCT diagnostics interpretable without significant training.

2.3.

Texture Analysis of OCT Images

Texture analysis is an image-processing technique that describes an image or portion of an image by characterizing its structure and pattern. Texture is made up of texture primitives. To describe texture it is necessary to describe the tone, or pixel intensity properties within the primitives, as well as the structure, which describes the spatial relationship between primitives. There are three principal approaches to texture analysis: statistical, structural, and spectral. Statistical analysis describes information about the content of the primitives, while structural describes the arrangement of the primitives, and spectral analysis describes the periodicity of the primitives.¹³

Gossage and Tkaczyk¹⁴ demonstrated that texture analysis can be used to analyze OCT images in an attempt to classify different tissue types. The study, which was able to distinguish between in vitro images of mouse skin (correct classification rate of 98.5%) and testicular fat (97.3%), as well as normal lung (88.6%) and abnormal lung (64.0%), analyzed the texture due to speckle in OCT images.

Although texture analysis has not been previously applied to the bladder, there have been studies that have used texture analysis to recognize dysplasia or cancer in other tissues. In two separate studies Qi ^{15, 16} applied texture analysis along with other image analysis techniques to OCT images of the esophagus to diagnose dysplasia. The resulting sensitivities were 87 and 82%, and the specificities were 69 and 74%. Likewise, image analysis was applied to the recognition of breast cancer by Zysk and Boppart.¹⁷ The results of the study indicated that their algorithm had a tumor tissue sensitivity of 97% and a specificity of 68%.

These studies indicate that texture analysis of OCT images is useful in distinguishing between types of tissue, but none of the studies involved images of the bladder. As far as we know, our study is the first in which texture analysis has been applied to OCT images of the lining of the urinary bladder.

3.1.

Image Preprocessing

Our algorithm was designed to characterize images based on a training set to identify features that will be useful in making a diagnosis. To remove as much variation as possible between the images, the dc bias and any existing background area were removed from the images before analysis.

Histogram analysis was used to identify the background in each image. Due to the low intensities associated with the lumen and bottom portion of the OCT image, the histograms produced a bimodal distribution, as seen in Fig. 2b . The first trough after the large low-intensity area was used to identify a threshold intensity to separate the background from the portion of the image containing information about the bladder lining. Once the threshold was determined, the portions of the image at the top and bottom of the image containing only intensities that were entirely below the threshold were removed from the image.

Fig. 2

To recognize the portion of the image (a) containing the most information, a smoothed histogram of the intensities in the image was created (b), and the image was thresholded (c).

Once the background was removed from the image, the mean intensity of the removed background was calculated. Since the background would theoretically be of regions in which there should be no signal (such as the lumen), or at the bottom of the image where little or no signal was measured, the mean intensity of the background would be a reliable estimate of the dc bias of the system. Consequently, the mean intensity of the removed background was used as an estimate of the dc bias and subtracted from the portion of the image containing information. Since the portion of the image containing information had higher intensities than the background, subtracting the mean intensity of the background did not and will not cause intensity levels to drop below zero.

3.2.

Texture Analysis

To develop our classification algorithm, we used five methods representing the three approaches to texture analysis (statistical, structural, and spectral) to obtain 74 texture features for each image in our training set.

The first method used cooccurrence matrices to analyze texture, which is a statistical method based on the repetitive nature of various intensity levels. The method requires that a neighbor be defined and a cooccurrence matrix created using that definition. The resulting cooccurrence matrix can be used to calculate six texture characteristics, where $P_{\phi ,d}(a,b)$ is the probability that a pixel with intensity $a$ will have a neighbor of intensity $b$ ; where neighbor is defined as the pixel at distance $d$ in the direction $\phi$ ; and ${\mu }_x$ , ${\mu }_y$ , ${\sigma }_x$ , and ${\sigma }_y$ are the means and standard deviations of the cooccurrence matrix¹³:

The second statistical method calculates a number of features based on the gray level histogram of the image.¹⁸ The first few features include the mean and second through fourth moments of the histogram. If the histogram is defined as $p\left(z_i\right)$ , where $z$ is the gray level, $L$ is the number of gray levels, and $i$ goes from 0 to $L-1$ , the remaining histogram features¹⁸ are ²

The third method used basic statistics to define four additional features for the image: the mean, standard deviation, range, and median.

The fourth method used a structural approach to texture analysis, Laws’ texture measures,¹⁹ to calculate 14 texture features. Laws’s method begins with three 1-D convolution kernels, representing averaging [ $L3=\left\{121\right\}$ ), edges ( $E3=\{-101$ ]), and spots ( $S3=\{-12-1\}$ ]), and eventually creates 14 texture features.

The final method used the spectral approach to texture analysis by taking the 2-D Fourier transform of the image and calculating the average energy values in several regions of the Fourier spectrum. We calculated the energy within six rings around the origin, within two horizontal bands close to the origin, and within two vertical bands close to the origin. We also calculated the normalized autocorrelation function, defined¹³ as $C_{\mathrm{ff}}=F^{-1}\{\mid F{\mid }^2\}$ , where $F$ represents the 2-D Fourier transform of the image. The resulting autocorrelation decreases slowly with increasing distance if the primitives in the texture are large, and decreases more quickly as the primitives decrease in size. We used the maximum autocorrelation value and the relative number of correlation values above three different thresholds as four additional texture features.

3.3.

Feature Selection

Once feature vectors were created for the training set, redundant features were removed, and the remaining features were compared for their ability to distinguish between images of non-cancerous and cancerous tissue.

To remove redundant features, we calculated the correlation matrix for the feature set, normalized it over the interval $[0,1]$ , and removed all but one feature from each set of highly correlated features (correlation $>0.85$ ). After the correlated features were removed, there were 18 remaining features in the feature vectors, which were then evaluated together and in groups for their ability to discriminate between tissue types. The 18 uncorrelated features are listed in Table 1 .

Table 1

Features remaining after removal of uncorrelated features.

3.3.1.

Scatter plots of features

After the feature set reduction, the resulting feature vectors were normalized over the interval [0, 255] and grouped according to their associated pathology results. Since research suggests that cancerous tissue disrupts the normally well structured, layered appearance of bladder tissue and appears more homogeneous than normal tissue, we hypothesized that the second moment of the histogram would be one of the strongest remaining features in distinguishing cancerous from noncancerous tissue, and used this as the second feature to produce 2-D plots for each pair of features.

The purpose of the plots was to identify whether the cancerous and noncancerous cases were indeed separable from one another, and whether they clustered tightly. Normal and exudative inflammation did tend to cluster together somewhat separately from dysplasia and CIS, which also tended to cluster together. Infiltrative inflammation, invasive tumors, and papillary lesions did not cluster well with themselves or each other. Figure 3 shows an example plot in which these generalizations are evident.

Fig. 3

Example scatter plot showing data set using two features: the second moment of the histogram and the maximum autocorrelation. For emphasis, a line was manually added to the plot to show the separability of dysplasia and CIS from normal tissue and exudative inflammation. This line was, however, not used to classify the data points. The line only serves to emphasize the separation evident in the image. All cases of dysplasia and CIS fall on one side of the line shown, while most cases of normal tissue and exudative inflammation fall on the other. The cases of infiltrative inflammation, papillary lesions, and invasive tumors are not separated by the two features.

Based on our observations from the feature plots, we excluded infiltrative inflammation from the training set for noncancerous tissue and began the decision process by assigning the data to one of two classes: normal/exudative inflammation and dysplasia/CIS. Each class then would be checked separately for instances of papillary growth and for invasion, as well as for infiltrative inflammation.

3.4.

Decision Tree

Based on the results of the feature selection process, a decision tree was created to classify any image into one of three classes: normal/exudative inflammation, dysplasia/CIS, or papillary $\mathrm{lesion}\pm \mathrm{invasion}$ . The decision tree shown in Fig. 4 uses various feature subsets at each decision point in the tree. The decisions are made using a discriminant function and the training data for the specific diagnoses included in each class. When distinguishing between cancerous and noncancerous tissue, the dysplasia/CIS and papillary $\mathrm{lesion}\pm \mathrm{invasion}$ classes are combined to form the more general cancerous class.

Fig. 4

Decision tree used in classification of images. Images are first subdivided into normal/exudative or dysplasia/CIS, after which each class is checked for cases of papillary $\mathrm{lesion}\pm \mathrm{invasion}$ , resulting in three possible end diagnoses: normal/exudative, dysplasia/CIS, or papillary $\mathrm{lesion}\pm \mathrm{invasion}$ .

Infiltrative inflammation was not included in the decision tree since none of the feature subsets were able to reliably distinguish infiltrative inflammation from either dysplasia/CIS or papillary $\mathrm{lesion}\pm \mathrm{invasion}$ . Cases of infiltrative inflammation that were diagnosed in the normal/exudative class were considered to be correct, so no attempt was made to distinguish between normal/exudative and infiltrative inflammation.

Once the decision tree was created, leave-one-out cross validation was used to determine the accuracy of the algorithm. Leave-one-out cross validation is a method of estimating classifier performance that does not require the data set to be divided into two separate training and test sets, but maintains independence between the test and training sets.²¹ The method requires that the following steps be repeated for each individual data point in the data set: remove the data point from the data set, use the remaining data set as the training data, test the removed sample using the mean vector and covariance matrix calculated from the training data, and return the removed data point to the data set. When this method is followed, each data point is tested on a training data set of which it was not a member, thereby maintaining the independence of the test and training data sets.

4.1.

Selected Features

While the trace values were useful when differentiating between subsets containing the same number of features, trace values tended to be higher when fewer features were in the subset. In all cases, a single feature was selected as the subset with the highest trace even though the selected feature never provided the best separation between classes. In general, however, the higher the trace value, the better the separation of the classes.

The first step in the decision tree separates the data into two classes based on the training data representing normal/exudative inflammation and dysplasia/CIS. The set of features selected for this step was the set of four features with the highest trace (0.98), and included the following features: (1) the correlation of the cooccurrence matrix with “neighbor” defined as the pixel immediately to the right, (2) the mean of the histogram, (3) the second moment of the histogram, and (4) the range of intensities in the image. When the discriminant function was used to separate the data using the preceding subset, 51 out of 54 normal/exudative images were classified correctly, and all images of dysplasia/CIS were classified correctly. While the set of four features was selected as the best, the subsets of two and three features, with trace values of 1.54 and 1.39, respectively, were also able to separate the classes effectively.

The second step of the decision tree consisted of two parts, one that attempted to distinguish images of papillary $\mathrm{lesion}\pm \mathrm{invasion}$ from normal/exudative, and one that attempted to distinguish them from dysplasia/CIS. None of the subsets had particularly high trace values when comparing normal/exudative images to papillary $\mathrm{lesion}\pm \mathrm{invasion}$ , indicating that it would be difficult to accurately separate the classes. The subset eventually chosen was the subset containing (1) the second moment of the histogram and (2) the maximum autocorrelation. The trace for this subset was only 0.59. When the discriminant function was used to classify the images, 47 out of the 51 normal/exudative images remained in the class normal/exudative while 10 of the 13 papillary $\mathrm{lesion}\pm \mathrm{invasion}$ images previously in the normal/exudative class were reassigned to the papillary $\mathrm{lesion}\pm \mathrm{invasion}$ class.

As with the previous comparison, none of the subsets had particularly high traces when comparing dysplasia/CIS with papillary $\mathrm{lesion}\pm \mathrm{invasion}$ . The subset chosen for this branch of the tree had a trace value of 1.08, and included the following three features: (1) the second moment of the histogram, (2) the third moment of the histogram, and (3) the amount of energy in a horizontal band around the origin of the Fourier transform of the image. The discriminant function left 10 of the 15 dysplasia/CIS images classified as dysplasia/CIS and moved 6 of the 11 papillary $\mathrm{lesion}\pm \mathrm{invasion}$ images previously in the dysplasia/CIS set to the correct class.

4.2.

Classification Results

Using the decision tree described previously, we classified (as either cancerous or noncancerous) each of the 182 images in the test set using leave-one-out cross validation. The results are shown in Table 2 . As expected from the scatter plots, the specificities for normal tissue and exudative inflammation are fairly high, 87 and 88% respectively, while the specificity for infiltrative inflammation is only 47%, giving an overall specificity of 62%. The sensitivities were 100, 93, and 78% for dysplasia and CIS, invasive tumors and papillary lesions, respectively.

Table 2

Algorithm performance: cancerous versus noncancerous.

The results for correctly distinguishing dysplasia/CIS and papillary $\mathrm{lesion}\pm \mathrm{invasion}$ are shown in Table 3 , along with the sensitivity for each specific diagnosis. The sensitivity for papillary lesions is 86%, but the others are lower, with dysplasia having a sensitivity of only 50%.

Table 3

Algorithm performance: dysplasia/CIS versus papillary lesion±invasion .

If the results of the first step in the decision tree are considered separately, it becomes apparent that the specificity was significantly reduced by the second set of decisions. The specificity after the first branch of the decision tree was 100% for normal images, 88% for exudative inflammation, and 80% for infiltrative images, producing an overall specificity of 85%. The sensitivity for the cancerous classes involved in the first step, dysplasia and CIS, was 100%. The implications of this observation are discussed later.