Keywords

1 Introduction

According to recent statistics published by the American Cancer Association, colorectal cancer is the second leading cause of deaths of men and women combined, with an estimated 101,420 new cases in 2019 (American Cancer [1, 9]). Early detection and removal of polyps significantly reduces the risk of death and screenings are recommended for those 50 or older. However, using clinical optical colonoscopy (OC), to screen the growing population of those over 50 is severely limited by our current resources. Alternatively, computed tomography colonography (CTC) has been a rising noninvasive solution to detect possible cancers [3, 5, 6, 11, 12], decreasing the amount of people being screened under OC. Furthermore, the majority of polyps are found to be nonneoplastic, named hyperplastic, which are abnormal growths with no risk, only taking up valuable resources when removed. Therefore, being able to distinguish between hyperplastic and adenomatous polyps through CTC screenings has become crucial.

In our previous research work, texture feature extraction techniques have been established and modified to improve the accuracy of diagnoses of polyps found through CTC screenings [4, 10]. Those texture features include intensity, gradient, and curvature. Pickhardt et al. suggests that the malignancy rates of polyps increase in correlation to their size [8]. Besides polyp’s malignancy rates, general kernel approach for gradient and curvature texture extraction may also get influenced by the polyp size [7]. Thus, in this study, we design experiments to investigate the effect of separating polyps based on size on the performance of machine learning.

The remainder of this chapter is organized as follows. Section 2 depicts our method for conducting feature analysis in computer-aided diagnosis of colorectal polyps. In Sect. 3, experimental design and results are reported. Finally, discussion of our work and conclusions are given in Sect. 4.

2 Materials and Methods

The aim of this study is to investigate the performance of machine learning on texture features based on different polyp sizes. Figure 1 shows a 6-mm small-sized polyp and a 22-mm medium-sized polyp visualized via two-dimensional (2D) axial image.

Fig. 1
figure 1

A 6-mm polyp (left) and a 20-mm polyp (right) on 2D axial image

Full size image

2.1 Flowchart of Our Method

The flowchart of our methods can be summarized in Fig. 2. First, from the original CT image, the volume of interest (VOI) of each polyp was extracted and further confirmed by the radiologists. Second, we applied an extended traditional Haralick model [2] with a total of 30 texture features [4], which were obtained from the VOIs of each polyp. In our study, the texture features of polyps consist of intensity, gradient, and curvature. Third, we performed a receiver operating characteristics (ROC)-based analysis to select the best feature sets. Finally, we employed a machine learning method, the Random Forest (RF) classifier, for feature classification.

Fig. 2
figure 2

Flowchart of our method for texture analysis

Full size image

2.2 Malignancy Risk of the Polyps

According to histopathology, polyps can be divided into two categories: nonneoplastic and neoplastic. Nonneoplastic polyps are benign polyps, including the subtypes of hyperplasic, mucosal polyp, and inflammatory, etc. On the other hand, neoplastic are malignancy risky, including serrated adenoma, tubular adenoma, tubulovillous adenoma, adenocarcinoma, etc. In this study, we labeled the nonneoplastic polyp as 0 and neoplastic polyps as 1, indicating their benign or malignancy risk accordingly.

2.3 Data Preparation

The dataset used in this study consists of a total number of 228 polyp masses found through a CTC database. Those polyp masses were confirmed by OC. The spatial resolution of the CT image is 0.7 by 0.7 by 1.0 mm3. All the polyp masses have a diameter size ranging from 6 to 30 mm. Polyps were grouped into three groups based on their sizes: 6–9 mm, 10–30 mm, and a combined group of 6–30 mm. For a fair performance comparison, we have a balanced polyp pool of 114 polyps in each size group. Within each size group, all of the datasets had 57 benign polyps and 57 malignant polyps.

3 Results

We investigated the performance of the algorithm on the 6–9 mm, 10–30 mm, and 6–30 mm groups. The studied features include intensity, gradient, curvature, all combined features. We generated a measure of area under the curve (AUC) values of ROC curve. The performance was determined by assessing the sensitivity and specificity values. The averaged AUC information is illustrated in Table 1. Figure 3 shows the averaged ROC curve for the intensity feature. Figure 4 shows the averaged ROC curve for the gradient feature. Figure 5 shows the averaged ROC curve for the curvature feature. And Fig. 6 shows the averaged ROC curve for all combined features.

Table 1 Averaged AUC information
Full size table
Fig. 3
figure 3

The averaged ROC curves for the intensity feature

Full size image
Fig. 4
figure 4

The averaged ROC curves for the gradient feature

Full size image
Fig. 5
figure 5

The averaged ROC curves for the curvature feature

Full size image
Fig. 6
figure 6

The averaged ROC curves for all combined features

Full size image

4 Discussion and Conclusions

Experimental results demonstrated that gradient and curvature were ideal distinguishing features for medium-sized polyps, whereas intensity was better for smaller-sized polyps. Due to their negligible proportions, the curvature and gradient features were flawed for the 6–9 mm polyps during the experiment. The opposite is true for the medium-sized polyp group.

When examining all features, the AUC value of the 10–30 mm polyps was greater than the AUC value of the 6–30 mm polyps, suggesting that separating the small- and medium-sized polyps would be beneficial in identifying medium-sized polyps. However, the AUC value of the 6–9 mm polyps was lower than that of the 6–30 mm polyps, suggesting that this separation may not be ideal for identifying smaller polyps. The AUC value for all polyps is greater than that of the individual group of smaller polyps because it is averaged out by the higher performance of the identification of the 10–30 mm polyps. Furthermore, the smaller polyps have less identifiable features. This study shall facilitate computer-aided diagnosis of polyps to achieve high performance by taking into account the contributions of different features among different polyp sizes.