Classification of human stomach cancer using morphological feature analysis from optical coherence tomography images

The light source is a MEMS-based wavelength swept source (HSL-20-100-B, of Santec Technologies, center wavelength: 1310 nm; main scan rate: 100 kHz; output power: 20 mW; tuning rang: 91.5 nm; coherence length: 13 mm). The interferometer system, including a balanced photodetector, is from Thorlabs (INT-MSI-1300B). The signal of the balanced photodetector is collected by a data acquisition device (ATS9350, Alarztec technologies, sampling rate: 500 M s⁻¹) and the fast Fourier transform is used to reconstruct the depth information of the tomographic images. This system has been described in our previous works [50–52] (figure 1).

Zoom In Zoom Out Reset image size

Each wavelength scanning cycle is an A-scan, providing only the image of one point position and the depth information. B-scans, comprising 512 A-scans in line, can image a section of each sample. Its image appears as a 2D tomographic image in the depth direction of one scanning line. Using multiple B-scans, scanning in repetition at the same region can reduce the random noise and enhance the signal-to-noise ratio in the average calculation, improving the image quality significantly overall. The axial resolution is approximately 12 µm in the air, the lateral resolution is approximately 22 µm, and the imaging depth is approximately 3 mm. The 2D scanning mirrors (X-Y plane) change the imaging angle, the X-axis mirror controls the B-scan, and the Y-axis mirror controls the en face scan.

All tissue samples were collected at the South China Hospital and South China Medical University (Guangzhou, China). Cancerous tissues were removed from patients during surgery. These tissues were tested with the OCT system immediately afterwards to ensure the integrity of the samples. We obtained 1539 OCT images of normal stomach tissue samples from six patients and 1567 OCT images of cancerous stomach tissue samples from eight patients. The scan range is approximately 5 mm  ×  5 mm and the imaging procedure is shown in figure 2. A photo of how OCT scanning is performed is shown in figure 2(a). An OCT B-scan image is shown in figure 2(b). A microscopic image of the stomach tissue under inspection is shown in figure 2(c) with the OCT imaging field marked in the box. The reconstructed en face image of the sample's surface using 3D volumetric data is shown in figure 2(d). The B-scan was along the long edge of the sample.

Zoom In Zoom Out Reset image size

After collecting OCT images with the SS-OCT system, the samples were prepared for standard pathological analysis. Biopsy samples were processed with the following procedures. Samples were fixed with 4% formalin solution, embedded in paraffin, sliced, and finally stained with hematoxylin and eosin. The histological sections are shown in figures 2(e) and (g). Sample preparation procedure led to sample tissues becoming dehydrated and deformed, which resulted in a relative distortion between the OCT images and histological images. A fresh strip of stomach sample was easy to twist and crimp. In order to keep samples flat when fixed and embedded, the samples were flattened on a small cover glass, inserted with a fine needle along the length of the samples to keep samples from distorting. The sample and whole cover glass could be immersed in formalin solution without shape distortions and provided a better surface during OCT scanning.

The first feature is the standard deviation at the 40th-pixel-depth line. The second feature is the standard deviation of intensity line of 0.25  ×  20th  +  0.5  ×  40th  +  0.25  ×  60th, and the third feature is the standard deviation of all 100 contour depth lines. The three additional quantitative morphological features represent the synthetical characteristics of the layered structures.

The OCT images of stomach normal tissue and cancerous tissue appeared remarkably different. As shown in figures 3(a) and 4(a), the OCT patterns of normal tissue are characterized by visibly smoother and homogeneous intensity distribution, while the cancerous tissue has a highly degenerated, fragmented, and heterogeneous texture. The low intensity feature spanning the top of these images was due to the impact of air, but there was some pattern noise in this area. The low intensity region at the bottom was a result from that limitation of the technique when penetrating into tissues. Therefore, there was little information in this range. The surface of the OCT image was not flat because the soft sample had a different thickness. We detected the surface and marked the boundary line as a white line. The range of 100 pixels (1.2 mm) in the depth direction was considered as ROI, which was the area between the two white lines, as shown in figures 3(b) and 4(b). The area of ROI could be flattened by eliminating the air area. The surface-flattened images are shown in figures 3(c) and 4(c).

Zoom In Zoom Out Reset image size

Zoom In Zoom Out Reset image size

According to the visual inspection, there was a significant change in texture observed between the OCT images for the normal and cancerous samples. In conventional image processing, the mean, the variance (σ², equation (1)), the third moment, and the entropy were used as parameters to describe the images' texture [43, 53]. The variance or standard deviation was the most significant criterion. We used the standard deviation (σ) of the whole ROI as a quantitative feature, which was the square root of the variance for the region [43] This parameter was named S1, which was the first quantitative feature.

$$\begin{align} \newcommand{\e}{{\rm e}} \displaystyle \sigma =\sqrt{{{\sigma }^{2}}}=\sqrt{\frac{1}{N}\sum\nolimits_{i=1}^{N}{{{({{x}_{i}}-\frac{1}{N}\sum\nolimits_{i=1}^{N}{{{x}_{i}}})}^{2}}}},\nonumber \end{align} \tag{ 1 }$$

where N is the pixels' amount, ${{x}_{i}}$ is the intensity of the ith pixel. The approach to using the standard deviation of the whole image was a global mean, but it cannot evaluate the textural information of a specific region. Local standard deviation for each pixel was used to express the degree of variation in a neighborhood of 11  ×  11 pixels. The value of each pixel's local standard deviation is shown as an image, as shown in figures 3(d) and 4(d). The local standard deviation was considered as the second quantitative feature named S2. The local standard deviation was calculated using the MATLAB function, 'stdfilt', the window size is 15 after many attempts. The texture variation of cancerous tissue was much more than that of normal tissue, so that the S1 and S2 of cancerous tissue should be larger than normal tissue.

Generally, normal stomach tissue was a regular structure, and each layer was homogeneous, but cancerous stomach tissue was heterogeneous. In the flattened images (figures 3(c) and 4(c)), the same depth of the surface-flattened image is equivalent for the same depth in real stomach tissue samples. As shown in figure 4(a), we take into consideration the imaging depth of the OCT (approximately 1.5 mm) and the thickness of the stomach tissue, and analyze the location cancerous tissue occurs (usually 0.3 mm below the surface). The major contribution of the morphological feature is usually located 0.3 mm below the surface. The contour depth line at the 20th, 40th, and 60th pixels of normal and cancerous tissues are shown in figures 5(a) and 6(a). The intensity of the three lines is shown in figures 5(b)–(d)) and 6(b)–(d).

Zoom In Zoom Out Reset image size

Zoom In Zoom Out Reset image size

Normal stomach tissue is continuous and the tissue's texture at the same depth is similar. The intensity of the three-contour depth lines of normal tissue was smooth, but that in the cancerous tissue appeared a dramatic variation. This pattern provided an available method to classify the cancerous tissue and can be quantified further from the standard deviation of the three lines. Three additional quantitative features were designed to describe one image. The standard deviation of the 40^th pixel depth line (figures 5(c) and 6(c)) is considered as the third quantitative feature, named S3. The standard deviation of 0.25  ×  20th  +  0.5  ×  40th  +  0.25  ×  60th (figures 5(e) and 6(e)) was considered as the fourth quantitative feature, named S4. Usually, the early-stage cancerous tissue appeared in the subsurface layer and then spread upwards and downwards. Therefore, the 40th-pixel depth line was given more weight. The standard deviation of the sum of all 100 contour depth lines (figures 5(f) and 6(f)) was considered as the fifth quantitative feature, named S5. Since the stomach tissue was a regular structure, the three contour lines represented the features of different layers. S3 showed the feature of a single layer, S4 represented the synthetical feature of three layers and S5 represented the synthetical feature of the all-100 lines.

Five quantitative parameters were extracted from each image and used as feature vectors to classify cancerous stomach tissue using five classifiers, i.e. SVM, KNN, RF, LR, and a typical threshold classification. For the classification of cancerous and normal tissue, a 'hold-out' validation technique was used to validate the classifiers. The holdout method divided the data into two mutually exclusive subsets: a training set and a testing set. The training set included a 500 normal tissue images and 500 cancerous tissue images, randomly selected, while the rest of the images were used as a testing set. Since the training set and testing set were separate, it avoided the overfitting and huge computation cost.

The automated classification algorithm was carried out with the MATLAB software. The SVM was realized by calling functions svmtrain and svmclassify, and the name-value pair arguments were set as 'kernel_function', 'quadratic'. For KNN, we used function 'knnclassify' with the nearest neighbor's value defined as 100. For RF, we used the function 'TreeBagger' to train the model with the parameter and value defined as 'OOBPred' and 'On'. The predict function was used to decide the testing set with the parameter defined as 'Trees'. For LR, the functions glmfit and glmval were used as the training and test procedures. Finally, the typical threshold classification analysis was conducted, where the images were labeled as cancerous tissue if the quantitative parameter exceeded the threshold. Otherwise, the image was defined as normal tissue.

The statistical results obeyed the following rules: a true negative means an image of normal stomach tissue was classified as normal, a false positive means an image of normal stomach tissue was wrongly classified as being cancerous, a true positive means an image of cancerous stomach tissue was classified as being cancerous, and a false negative means an image of cancerous stomach tissue was labeled as being non-cancerous (normal).

$$\begin{align} \newcommand{\e}{{\rm e}} \displaystyle {\rm Sensitivtity}=\frac{{\rm TP}}{{\rm TP}+{\rm FN}}\nonumber \end{align} \tag{ 2 }$$

$$\begin{align} \newcommand{\e}{{\rm e}} \displaystyle {\rm Specificity}=\frac{{\rm TN}}{{\rm TN}+{\rm FP}}\nonumber \end{align} \tag{ 3 }$$

$$\begin{align} \newcommand{\e}{{\rm e}} \displaystyle {\rm Accuracy}=\frac{{\rm TP}+{\rm TN}}{{\rm TP}+{\rm FN}+{\rm TN}+{\rm FP}}.\nonumber \end{align} \tag{ 4 }$$

The classification sensitivity (equation (2)), specificity (equation (3)), and accuracy (equation (4)) were used to evaluate these classifiers and compared with each other. With the typical threshold classification method, if the threshold value gradually increased, the ROC curve can be drawn out with the sensitivity and 1-specificity.

Figure 7 shows the boxplot and p  values of the five quantitative features of the OCT images of cancerous stomach tissues in the S1, S2, S3, S4 and S5. The five quantitative features in images of cancerous tissue were distinguishable from that of images bearing normal tissue. It can be observed from figures 7(a)–(e) that the five quantitative features have significant differences in their mean intensity value of cancerous tissue compared to normal tissue.

Zoom In Zoom Out Reset image size

To visualize the effect of the SVM identification, we used the two features, S3 and S4, to identify the OCT images. Figure 8(a) shows the classified procedure of SVM with the features of S3 and S4, and the enlarge region (figure 8(b)) demonstrated the support vectors' data points as outlined circles. Support vectors' data points were based on the training set, and the black line was the classified parameter based on the support vectors' data points, which was used to decide the testing dataset.

Zoom In Zoom Out Reset image size

The SVM was employed to classify all five quantitative features, and the result was listed in table 1. The classification of S4 presented the best result among all five quantitative features; the sensitivity, specificity and accuracy are 95.34%, 97.21% and 96.27%, respectively. The classified effect based on the local image information was a little better than the general whole image information. The S3, S4 and S5 were specially designed for gastric cancerous tissue based on its regular structure, and the classification result of them was better than the general morphological analysis (S1 and S2).

The result of KNN, LR, RF, and threshold classification were listed in the following (tables 2–5), and they appeared with similar rules to the SVM. From the result of tables 1–5, the RF method appears to be the best result and the LR method presents the worst result. There is not enough statistical evidence demonstrating that the RF method is the better algorithm than the other four to classify the stomach cancerous OCT images. All five algorithm get very excellent results; the difference between these means was tiny. All five-classification results showed that S3, S4 and S5 are remarkably better than S1 and S2. Hence, the designed feature based on the layer structure of stomach tissue was significantly more effective.

The ROC curve of the threshold classification under five features is shown in figure 9. Each curve reflects the rates at which false and true positive values are determined at different thresholds. S4 had the best performance in this test, S3 shared a similar result with S5. S1 and S2 were worse than the previous three parameters. The AUC of the ROC curve was then utilized to estimate the classification method (table 6). The AUC of S4 was up to 99.65%. The ROC curve demonstrates that S3, S4 and S5 have a better performance than S1 and S2.

Zoom In Zoom Out Reset image size