Developing a New Radiomics-based CT Image Marker to Detect Lymph Node Metastasis Among the Cervical Cancer Patients

Abstract

Background and Objective:

In diagnosis of cervical cancer patients, lymph node (LN) metastasis is a highly important indicator for the following treatment management. Although CT/PET (i.e., computed tomography/positron emission tomography) examination is the most effective approach for this detection, it is limited by the high cost and low accessibility, especially for the rural areas in U.S.A. or other developing countries. To address this challenge, this investigation aims to develop and test a novel radiomics-based CT image marker to detect lymph node metastasis for cervical cancer patients.

Methods:

A total of 1,763 radiomics features were first computed from the segmented primary cervical tumor depicted on one CT image with the maximal tumor region. Next, a principal component analysis algorithm was applied on the initial feature pool to determine an optimal feature cluster. Then, based on this optimal cluster, the prediction models (i.e. logistic regression or support vector machine) were trained and optimized to generate an image marker to detect LN metastasis. In this study, a retrospective dataset containing 127 cervical cancer patients were established to build and test the model. The model was trained using a leave-one-case-out (LOCO) cross-validation strategy and image marker performance was evaluated using the area under receiver operation characteristic (ROC) curve (AUC).

Results:

The result shows that the SVM-generated imaging marker achieved higher performance with AUC value of 0.841 ± 0.035. When setting an operating threshold of 0.5 on model-generated prediction scores, the imaging marker yielded a positive and negative predictive value (PPV and NPV) of 0.762 and 0.765 respectively, while the total accuracy is 76.4%.

Conclusions:

This study initially verified the feasibility of utilizing CT image and radiomics technology to develop a low-cost image marker to detect LN metastasis for assisting stratification of cervical cancer patients.

1. INTRODUCTION

In gynecologic oncology, cervical cancer is the third most prevalent carcinoma, which particularly affects the young females aged from 20 to 39 years [1]. In United States, it may approximately affect 13,800 women and account for 4,290 deaths in 2020 [1]. For the patients with early stage cervical cancer, radical surgery is suggested as the most effective treatment [2]. However, for the women with more advanced diseases, the effectiveness of surgery is limited, and the curative treatment is concurrent chemoradiotherapy [3]. Therefore, it is critically important to categorize the cervical cancer patients for achieving the optimal treatment efficacy.

In clinical practice, the most critical indicator for stratifying the cervical cancer patients is based on the status of the lymph node metastasis [4]. Currently, lymphadenectomy and sentinel node biopsy are the most accurate approaches to determine metastasis, as the sampled lymph node tissues will directly indicate the tumor spread under the following pathological examination. However, both of these two approaches are invasive examinations, which may cause additional pain and complications on patients. Thus, in order to avoid lymphadenectomy or biopsies, a non-invasive imaging examination method is clinically needed to detect the lymph node metastasis for treatment management. Among different types of imaging methods, CT/PET (computed tomography/positron emission tomography) examination is considered the most effective method and has been used in many academic medical centers as the standard procedure. For this method, PET indicates the metabolic activity of the target lymph node, while CT can provide the location and anatomical information. Despite the effectiveness, PET is a nuclear medicine based imaging modality, which requires much more complicated nuclear imaging agents and operation facility. Therefore, CT/PET examination is a long time procedure with high cost, which causes heavy financial burden on cancer patients. In addition, the high cost and complicated operation procedure of the CT/PET imaging modality also limits its accessibility to many patients, especially in the rural areas of U.S.A. or other developing countries.

As compared to CT/PET examination, CT imaging alone has the advantages of low operating cost and wide accessibility. However, the CT image interpretation is visually conducted by the radiologists, and the CT image based LN metastasis identification is not as accurate as CT/PET examination [5]. In order to address this clinical challenge, it is crucial to identify new quantitative image markers to improve the accuracy of CT image based LN metastasis detection. For this purpose, we recognize that radiomics is an emerging technology that receives intensive research focus during the last ten years to search for new quantitative image markers [6]. This new technique extracts a large amount of quantitative image features from a variety of imaging modalities (e.g. CT, MRI, etc.), based on which an image marker is generated using a machine learning based prediction model. For instance, one study extracted a total of 440 features to quantify the tumor phenotype from 1,019 lung cancer or head & neck cancer patients. Using radiomics analysis, four features were finally selected as a prognostic signature, which shows superior performance in patient survival prediction as compared to the conventional method (i.e. TNM staging) [7]. The second study utilized a total of 388 quantitative features to analyze the MRI images from 365 glioblastoma patients, and a 24 feature cluster were finally determined to generate an image marker to stratify the patients into three different survival groups (i.e. poor, intermediate, and favorable groups) [8]. Beside the survival analysis, radiomics based image marker has also been successfully utilized in many cancer related applications including the chemotherapy response prediction [9] or benign / malignant lesions classification [10–13]. However, to the best of our knowledge, no studies have focused on applying radiomics method to develop a CT image marker to identify the LN metastasis for cervical cancer patients.

Thus, in this study, we hypothesized that the clinical information (e.g. relate to shape, density, texture, etc.) contained inside the primary cervical tumor are highly associated with underlying cell biology mechanisms (e.g. cycling pathways of tumor cell), which are also closely connected to the lymph node metastasis (LNM). Thus, we can utilize the novel radiomics technology to analyze the primary tumor and uncover this useful information, which will generate a CT image marker for effectively detecting LN metastasis. To test this hypothesis, we assembled a retrospective image dataset, and applied radiomics concept to create an initial feature pool involving a comprehensive list of 1,763 handcrafted features, which were computed from the primary cervical tumor regions depicted on CT images. Then, an optimal and low-dimensional feature vector was identified and generated by a principal component analysis (PCA) algorithm, which were utilized as the input of a logistic regression model to generate a new CT image marker for detecting LN metastasis. The details of this study are presented as follows.

2. MATERIALS AND METHODS

2.1. Database

In this study, the CT dataset was retrospectively collected from our medical center, which consist of 127 locally advanced cervical cancer patients, which were identified based on the following two inclusion criteria: 1) The patients were diagnosed of stage IIb-IVA cervical cancer of all histology types (i.e. squamous cell carcinoma, adenocarcinoma or adenosquamous carcinoma); and 2) They received systematic concurrent chemoradiotherapy in the cancer center of our university. For each patient, we collected the pre-therapy CT images, which were obtained using GE LightSpeed VCT 64-detector or GE Discovery 600 16-detector CT machines with a standardized image acquisition protocol previously established at our medical center. The protocol is briefly described as follows: (1) X-ray power output is set at 120 kVp and a variable range from 100 to 600mA depending on patient body size. (2) The 100cc contrast agent of Isovue 370 is intravenously injected using a standard power injector with a rate at 2–3cc/sec through a 20-gauge IV needle placed at the antecubital fossa. (3) Two phases of CT scans are performed in each examination. The 1st scan phase begins 60 seconds after start of contrast injection and the 2nd (delay) phase begins 5 minutes after contrast injection. Each phase of scan takes ~4 seconds. (4) CT image scans at 5mm with an axial reconstruction to 1.25 mm, sagittal and coronal reconstructions to 2.5 mm. Among these 127 patients, 52 exhibited lymph node (LN) metastasis and the other 75 did not have LN metastasis.

2.2. Image Feature Computation

Before feature computation, the primary cervical tumor was visually segmented by the experienced researchers and radiologists. All the tumor segmentation and feature computation were conducted on the images reconstructed from the first phase of the CT scan. Since a tumor is typically depicted on a series of CT slices, some researchers tried to compute radiomics features from multiple CT slices that depict tumor regions [7], which are defined as 3D features. Theoretically, the 3D feature should be more accurate, as it carries more information to characterize the tumor heterogeneity. However, the previous study found no significant performance difference between using 2D and 3D radiomics feature based markers [14], which can be attributed by the following factors. First, the 3D feature must be conducted based on 3D tumor segmentation, and multiple slice segmentation will enlarge the segmentation errors multiple times, which counteracts the advantages of 3D feature computation and reduces the feature robustness. Second, the 3D tumor accuracy is heavily dependent on the axial scanning resolution of the CT images. The diversified scanning protocol is unavoidable in multiple site studies. Therefore, 3D tumor features may have high generalization errors than 2D features. Meanwhile, 2D features can have the advantages of low computation complexity and wide availability, as most of the image features are based on 2D computation. Accordingly, a total of 1,763 quantitative image features were extracted on each segmented tumor region, which can be divided into the following 3 groups: 1) Shape and density features; 2) Histogram based features; and 3) Multiscale features. All these image features characterize tumor heterogeneity in both time and frequency domains.

2.2.1. Shape and Density Feature Group

As illustrated in Table 1, this group includes a total of 15 features to describe the shape and density properties of tumor region. The pool of shape features includes one feature to compute the tumor area, 8 features to describe the tumor geometric irregularity, and two features to describe the tumor complexity. These irregularity features are based on shape ratio (e.g. ratio of area to perimeter length), tumor radial length (e.g. Shannon entropy of radial length), or central position shift. Specifically, the central position shift feature estimates the distance between pixel with smallest density (i.e. grayscale intensity) and the gravity center of the tumor region [15].

Table 1:

A list of computed shape and density features

Class	Description
Shape features	Tumor area, ratio of area to perimeter (roundness), ratio of outer rectangle height to width, ratio of tumor region maximal radius to minimal radius, skewness of radial length of tumor region, normalized mean radial length, standard deviation of radial length, Shannon entropy of radial length, normalized central position shift
Density features	Average pixel value in the tumor area, standard deviation of pixel values in the tumor area, region contrast, Shannon entropy of pixels values in tumor region, fractal dimension, $C_0$ complexity

Tumor complexity is computed by two features: $C_0$ complexity and fractal dimension. $C_0$ complexity is computed as follows [16]: First, given a $N\times N$ gray-level image $\left\{f\left(j,k\right),j=\mathrm{0,1},\dots ,N-1,k=0,1,\dots ,N-1\right\}$, its 2-dimension discrete Fourier transform (DFT) is represented by $\left\{F\left(m,n\right),m=\mathrm{0,1},\dots ,N-1,n=0,1,\dots ,N-1\right\}$. In DFT domain, we keep all the elements with values higher than the mean square value:

$$\tilde{F}(m,n)=\left\{\begin{array}{c}F\left(m,n\right),if{\left|F\right(m,n\left)\right|}^2>G\\ 0,if{\left|F\right(m,n\left)\right|}^2\le G\end{array}\right.$$ 1)

where G is the mean square of F(m, n): $G=\frac{1}{N^2}\sum _{m=0}^{N-1}\sum _{n=0}^{N-1}{\left|F\left(m,n\right)\right|}^2$. Next, we compute the inverse discrete Fourier transform $\tilde{f}\left(j,k\right)$ of $\tilde{F}$, and the complexity $C_0$ is defined as:

$$C_0=\sum _{j=0}^{N-1}\sum _{k=0}^{N-1}{|f\left(j,k\right)-\tilde{f}\left(j,k\right)|}^2/\sum _{j=0}^{N-1}\sum _{k=0}^{N-1}{\left|f\left(j,k\right)\right|}^2$$ 2)

Besides $C_0$ complexity, fractal dimension [17] is another feature to characterize the tumor complexity [18]. For this feature, the target $N\times N$ image is first down-sampled to a $s\times s$ image, and the ratio of s to n is defined as down sampling scale: r = s/n. Next, the original image is divided into a number of sub-regions, each of which has a size of $s\times s$. For the subregion (i, j), $n_r\left(i,j\right)=l-k+1$ if the maximum and minimum intensity value falls into the lth and kth box. Let $N_r=\sum n_r\left(i,j\right)$. Then, the fractal dimension of $S$ is given as follows:

$$FD=\frac{N_r}{\mathrm{l}\mathrm{o}\mathrm{g}\left(\raisebox{1ex}{$1$}\!\left/ \!\raisebox{-1ex}{$r$}\right.\right)}$$ 3)

In addition to the shape features, density features characterize the tumor heterogeneity by computing the statistical measures of the CT number (Hounsfield unit). within the tumor area, which includes average pixel value, standard deviation of pixels values (coherence), region contrast and Shannon entropy.

2.2.2. Histogram Based Feature Group

a). Local Binary Pattern Feature

For this type of the feature, we first extract the eight neighbors for each pixel of the target image, and then compute the average intensity of these 9 pixels (target pixel and its eight neighbors) [19]. Next, each of the eight neighbor pixels is compared with the average intensity, and 1 or 0 will be assigned to the neighbor if its intensity is higher or lower than the average. As illustrated in Figure 1, these eight neighbor binary sequences generate a one-byte value ranging from 0 to 255, which will be assigned as the pattern value for the target pixel. After applying this procedure for all the pixels of the target image, a pattern image is created, for which each pixel value is the above pattern value. The histogram of this pattern image is used as a 256-value feature vector in this study.

Figure 1.

The calculation of local binary pattern feature

b). Edge Histogram Descriptor

The edge histogram descriptor (EHD) is a useful texture signature that characterizes the spatial distribution of image edges [20], which is computed as follows. The original image is first evenly divided into 16 sub-images, each of which is then further partitioned into a certain number of image blocks. For each image block, five edge detectors are applied to detect vertical, horizontal, 45° diagonal, 135° diagonal, and isotropic edges, respectively. The image block is labeled as containing edge with one of the above orientations if the convolution between the block and the corresponding edge detectors is above the preset threshold. For each sub-image, the edges with five different orientations is computed as a histogram. Accordingly, for an image with 16 sub-images, the EHD feature is computed as a feature vector with a dimension of 80.

c). Histogram of Oriented Gradients

The histogram of oriented gradients (HOG) is a widely used feature and the central idea is that the distribution of intensity gradients can capture the appearance and shape of local objects within an image [21]. The HOG feature extraction can be summarized as follows. First, the values of image pixels should be normalized using gamma transformation or histogram equalization. Next, the gradients at different orientations are computed. Similar to the EHD feature computation, the image is divided into a total of M blocks, while each block is divided into N cells. For each cell, the first-order differential calculation is performed along K different orientations, which are evenly ranging from 0~180 degrees, and a histogram is generated against the gradient direction. Then, the HOG feature vectors are normalized within blocks using L2 norm. After combining the HOG of all the cells together, a final feature with K × M × N dimension is generated, where K represents the number of direction angles in each cell, M and N represent the number of blocks and the number of cells in a block, respectively. In this study, the K, M, and N are determined as 9, 20 and 4, which has been verified as an optimal balance between the feature performance and complexity.

2.2.3. Multiscale Feature Group

a). Pyramid and Wavelet Based Features

To compute this type of the feature, the original image is first decomposed using Laplacian pyramid or wavelet methods. For Laplacian pyramid method, the procedure is as follows [22]: A Gaussian kernel based low pass filter is applied on the original image, and the difference between the original and low pass filtered images is considered as the level 1 decomposition (L1). Then, the L1 image is down sampled by a scale of 2, on which the level 2 decomposition is generated using the same approach. L3 decomposition is computed similarly, and L1-L3 decompositions are used for the following feature computation.

Meanwhile, in Wavelet decomposition, the Harr wavelet transform [23] is first applied on the original image, and the resulting two-dimensional array of coefficients contains four bands of data, which are labelled as LL (low-low frequency component), HL (high-low frequency component), LH (low-high frequency component) and HH (high-high frequency component) respectively. The LL band captures the low frequency components (smooth variations) that constitute the base of an image, the HH band retains high frequency components (edges details) that can refine the image. The LL band can be further decomposed in the same manner, thereby producing more sub-bands. Similar to pyramid decomposition, L1-L3 wavelet decompositions are conducted and used in this study.

As illustrated in Figure 2, each decomposition is evenly divided into four regions namely, A, B, C, D. The center of region E is at the central position of the original image, and it has the same size as regions A-D. Three features are computed on original region and divided regions A-E, including 1D complexity, 2D complexity and FD. For 1D decomposition, all the pixels within the region is reorganized into a 1D array, on which the complexity is computed. Hence, a total of 108 features are generated for all Laplacian pyramid or wavelet decompositions.

Figure 2.

Subregion partition of the sub-band LL. The entire band is evenly divided into four regions (A-D). Region E locates in the band center, with the same size as regions A-D.

b). Gabor Features

The frequency domain features contain copious clinically meaningful information for patient stratification [9]. Among different types of frequency domain features, Gabor based wavelet transform are optimal for measuring local spatial frequencies at different scales and orientations [24]. The Gabor transform can be defined as follows [25]:

$$\mathrm{G}\left(\mathrm{x},\mathrm{y}\right)=\left(\frac{1}{2\pi {\sigma }_x{\sigma }_y}\right)\mathrm{e}\mathrm{x}\mathrm{p}(-\frac{1}{2}(\frac{x^2}{{\sigma }_x^2}+\frac{y^2}{{\sigma }_y^2})+2\pi Wx)$$ 4)

In this study, we utilize a number of self-similar functions to decompose the original image, which is implemented by rotating and dilating the mother Gabor function $\mathrm{G}\left(\mathrm{x},\mathrm{y}\right)$ as follows:

$${\mathrm{g}}_{mn}\left(\mathrm{x},\mathrm{y}\right)=a^{-m}\mathrm{G}\left(\mathrm{x}\mathrm{\text{'}},\mathrm{y}\mathrm{\text{'}}\right)$$ 5)

where ${\mathrm{x}}^{\mathrm{\text{'}}}=a^{-m}\left(\mathrm{x}\cos \theta +\mathrm{y}\sin \theta \right)$, and $\mathrm{y}\mathrm{\text{'}}=a^{-m}\left(-\mathrm{x}\cos \theta +\mathrm{y}\sin \theta \right)$. The parameter $\theta =\frac{n\pi }{n}$ defines the orientation and scale of Gabor filters. For our application, there are 3 scales (m = 3) and 4 orientations (n = 4), resulting in 12 filters. For each filtered image, we calculate the mean and standard deviation. Thus, a 24-dimensional Gabor feature vector can be obtained.

Similarly, using Harr wavelet transform [26], we decompose the original image into different scales. At scale L1-L3, we still estimate the mean and standard deviation (STD) on each of the four sub-bands (i.e. LL, LH, HL, HH bands). Therefore, a total of 24 features are computed. Furthermore, we use similar method to generate another 24 features for Daubechies 2 wavelet (DB2) transform [27].

c). GIST Descriptor

GIST characterizes a set of perceptual properties closely related to the dominant target structure including naturalness, openness, roughness, expansion, ruggedness [28]. Specifically, GIST captures the gradient information at different scales and orientations for different parts of an image. For this purpose, a total of 32 Gabor filters at 4 scales and 8 orientations are first applied on the given image to generate 32 feature maps. Each feature map is then divided into 4×4 equal regions. On each region the filter coefficients are averaged as one feature number. As a result, each feature map creates 16 feature values and a total of 512 features are finally computed by the 32 feature maps as the GIST descriptor.

2.3. Develop a Machine Learning based Model to Predict Lymph Node Metastasis

After computing all features discussed in above sections, we developed a multi-feature fusion based machine learning model to predict LN metastasis. Although many different machine learning models can be used to predict risk of LN metastasis, we selected to build two different machine learning models namely, a logistic regression (LR) [29] based and a support vector machine (SVM) [30] based prediction model, which is due to the limited size of our dataset of 127 cases. The model generates a prediction score indicating the estimated risk or likelihood of LN metastasis for an individual patient. Among different types of the machine learning based classification models, logistic regression has relatively simple architecture with strong robustness, which is easy to be optimized with high performance while avoiding the model overfitting. Given that the number of features (i.e. 1763) is much larger than the number of cases and the information redundancy cannot be avoided in the original feature set, dimension reduction is first applied on the initial feature pool to create an uncorrelated optimal feature cluster. In this study, the principle component analysis [31] is adopted to reduce the feature space dimensionality and redundancy. For this method, let X be the feature matrix, for which each column represents one type of feature and each row represents one sample. Next, the sample co-variance matrix X^TX is computed and defined as M, on which the characteristic transform is performed: M = WΣW^T. All the characteristic vectors and values are arranged with decreasing order in eigenvector matrix W and diagonal eigenvalue matrix Σ. Since the eigenvalues decrease rapidly, only a small number of coordinates will be able to accurately approximate the original data matrix. In this study, we only keep the first 18 components. These components will be used to build the prediction model. As a comparison, we also directly apply each class of the original features to build the model for metastasis prediction.

Next, the leave-one-case-out (LOCO) cross-validation technique is adopted to train the logistic regression model and evaluate the model performance. LOCO is essentially a special form of k-fold cross-validation, where the number of folds is equal to the number of training instances (e.g., n). As compared to the regular K-fold (e.g. K = 5) cross validation, the LOCO strategy has the least bias to optimize the prediction model, because it eliminates the potential case partition bias and increases the size of training samples to the maximum within the available dataset. Therefore, the LOCO is well-recognized as the optimal option to train and test machine learning models when the total dataset is limited [32]. Finally, the trained logistic regression model generates a prediction likelihood score ranging from 0 to 1, in which the higher score indicates the higher likelihood of LN metastasis. The model performance is assessed using the receiver operating characteristic (ROC) curve. All results are tabulated for performance comparison and analysis in which statistical significance (p-value) is also computed.

3. RESULTS

Figure 3 (a) is the heatmap of all the 1,763 features, which is generated by a total of 127 observations. In this map, each entry represents the colorized correlation coefficient ranging from 0 (Blue) to 1 (red). The map shows that the group of LBP features has less independencies than the others. All these coefficients are categorized into six evenly distributed intervals between 0 and 1, and the results are illustrated as a histogram (Figure 3 (b)). The histogram reveals that more than 60% of the correlation coefficients falls into the interval ranging from 0–0.4, implying that our initial feature pool covers comprehensive tumor characteristics with small information redundancy.

Figure 3.

a) Heatmap of all the 1763 features b) Histogram of the feature correlation coefficients

After applying the PCA algorithm, a total of 18 principal components (PC) are selected. Given that each component is a linear combination of all the original features, PCA coefficients can directly reflect the impact of original image features on each individual component. To investigate this impact, the coefficient distribution is generated for each component. Accordingly, the coefficients belonging to one feature group are added together, and the summed results are then normalized to create the histogram. These distribution histograms were organized as an impact heatmap in Figure 4(a). In this map, horizontal and vertical direction represent feature group and principal components, respectively. Hence, each column represents the normalized coefficient distribution of one component. The map demonstrates that the Wavelet-DB2 feature group has stronger impact than the other features. For example, in component 3, the Wavelet-DB2 coefficient is 1.0, which is much larger than the second (Shape and Density, 0.514) and third (LBP, 0.368) most important group. Meanwhile, the coefficients of seven feature groups (i.e. HOG, EHD, GIST, Wavelet-1D-Complexity, Pyramid-1D-Complexity, Pyramid-2D-Complexity, and Pyramid-FD) are lower than 0.1 for most of the components, which shows that these features do not significantly contribute to the components generation. In addition, the discriminative power of each individual PC was assessed using area under the ROC curve (AUC), which are sorted in decreasing order ranging from 0.511 to 0.744 (Figure 4(b)).

Figure 4.

a) Feature coefficients for each principal component b) The AUC values of the principal components. All the components were sorted with decreasing order.

Using these PCs as input, two different models (i.e. logistic regression and support vector machine) were trained and optimized to predict whether the patient has LN metastasis or not. The model performance was assessed using ROC curve, which was generated by maximum likelihood based curve fitting algorithm (ROCKIT, http://metz-roc.uchicago.edu/, University of Chicago). As demonstrated in Figure 5, the SVM model yields the AUC (i.e. area under the ROC curve) value of 0.841± 0.035, while the AUC value yielded by LR model is 0.814 ± 0.037, which is slightly lower than the SVM model without a statistically significant difference ( p = 0.075). Thus, two models can indicate higher discriminative power in LN metastasis prediction. As a comparison, we also utilized the best and second best performed PC to achieve the prediction, while AUC values of 0.710 ± 0.047, 0.662 ± 0.047 were yielded, respectively. Table 2 (a) lists the confusion matrix of the prediction results of the SVM prediction model. When setting the operating threshold of 0.5 on the model-generated likelihood score, this model grouped 42 and 85 cases as metastasis “positive” and “negative,” respectively. Among the “positive” prediction group of 42 cases, a total of 32 were confirmed as positive by CT/PET examinations, achieving a positive prediction value (PPV) of 0.762. Meanwhile, 59 out of the 78 cases in the “negative” class was confirmed by CT/PET examinations, and the corresponding negative prediction value (NPV) of 0.765. Combining both groups together, the total prediction accuracy is 76.4% (97/127). Similarly, for the LR model, the PPV, NPV, and total predicting accuracy are 0.673 (33/49), 0.756 (59/78), and 72.4% (92/127), respectively, which are illustrated in Table 2 (b).

Figure 5.

ROC curve of LN metastasis prediction results

Table 2:

Confusion matrix of LN metastasis prediction

(a) SVM model
	Predicted: positive	Predicted: negative
Actual: positive	32	20
Actual: negative	10	65
Accuracy	Positive prediction value	Negative prediction value
0.764	0.762	0.765

(b) LR model
	Predicted: positive	Predicted: negative
Actual: positive	33	19
Actual: negative	16	59
Accuracy	Positive prediction value	Negative prediction value
0.724	0.673	0.756

As a comparison, we also used each group of the features as the input to optimize the regression model for LN metastasis prediction, and all the results were summaried in Table 3. It is revealed that the AUC values ranges from 0.532 to 0.740. In five out of the fourteen groups, the AUC values is higher than 0.65, indicating that each group of image features have significantly higher dicriminative power than random guess (AUC = 0.5) in performing this prediction task. The value of 0.65 is used as a threshold is because that most previously reported cancer prognostic markers have an minimum AUC of 0.6 [33–37]. Specially, the Shape & Density and Wavelet-DB2 feature group achieved the best and second best performance, with AUC values of 0.740±0.044 and 0.737±0.045, respectively. The results indicate that the well trained logistic regression or SVM model yields significantly higher performance (AUC value) than that of the two above single optimalfeatures (e.g. Shape and Density, Wavelet-DB2). For instance, when comparing the LM model with these two single optimal features, the statisitcally significant differences were detected with p values of 0.024 and 0.026, respectively.

Table 3:

Comparison of LN metastasis prediction performance of the fusesd feature and seperate feature groups

Feature type	AUC ± STH
Fused feature (SVM)	0.841 ± 0.035
Fused feature (LM)	0.814 ± 0.037
Shape and Density	0.740 ± 0.044
Wavelet-DB2	0.737 ± 0.045
Pyramid-FD	0.694 ± 0.049
Wavelet-FD	0.681 ± 0.051
Wavelet-Haar	0.673 ± 0.047
Gabor	0.646 ± 0.054
EHD	0.621 ± 0.053
Pyramid-1D-Complexity	0.607 ± 0.049
Pyramid-2D-Complexity	0.603 ± 0.050
Wavelet-2D-Complexity	0.589 ± 0.050
Wavelet-1D-Complexity	0.555 ± 0.051
HOG	0.539 ± 0.057
Gist	0.533 ± 0.061
LBP	0.514 ± 0.059

4. DISCUSSION

In this study, we developed and evaluated a novel and cost-effective image marker to predict LN metastasis for locally advanced cervical cancer patients. This study has several unique characteristics as follows. First, although radiomics features have been utilized to predict patient prognosis of many different cancers [7, 38–40], to the best of our knowledge, this is the first study that applies the CT image based radiomics features to predict the LN metastasis of cervical cancer patients. Currently, the standard update value (SUV) of the PET images is the most important criterion for diagnosis of LN metastasis, as the metastatic tumor activity is mainly indicted by the glucose consumption, which can be quantified by PET examination using FDG tracer (2-fluoro-2-deoxy-D-glucose). However, the model proposed in this study identified and computed a large amount of radiomic features from the primary cervical tumor depicted on the CT images. After applying the PCA algorithm to generate an optimal feature vector, a machine learning based model was optimized for predicting the LN metastasis. Using a retrospectively assembled clinical image dataset, the experimental result demonstrated that the optimized model yielded an AUC value of 0.814 ± 0.037, which implies that the valuable and discriminative metastasis information or features are also contained within CT images. As compared to PET/CT examination, this CT image marker has many potential advantages of low cost, short examination time, simple imaging scanning procedure and wide accessibility.

Second, we computed and assembled a comprehensive feature pool with 1,763 features, which can be grouped into 3 major categories (14 subcategories). Considering the number of feature categories and the quantity of features, we believe that this is by far the most comprehensive feature pool designed specifically for the task of predicting LN metastasis based upon CT images. The results demonstrate that five out of the 14 groups (i.e. Shape and Density, Wavelet-DB2, Pyramid-FD, Wavelet-FD, and Wavelet-Haar) contain significant discriminative power in classifying LN metastasis cases (i.e. AUC of 0.65 or higher), which implies the effectiveness of the initial feature pool. Although the Shape and Density group feature achieved the highest prediction performance, three of the five best performed feature groups are based on wavelet transform, which demonstrates that the frequency domain information is equally important as compared to the spatial domain. Meanwhile, two FD based feature categories yield AUC values of 0.694 and 0.681, respectively, which may reveal that the tumor pattern complexity should be an effective indicator to manifest the lymph node metastasis. Additionally, four groups of the features were performed at different scales of the segmented tumors, which implies that the clinically meaningful metastasis information may also be depicted in the detailed tumor patterns.

Third, a novel machine learning based image marker was generated by fusing 18 non-redundant radiomics feature components, which were determined by principal component analysis algorithm. As shown in Table 3, the fused feature gains 7% AUC value increase as compared to the separate feature groups. The performance improvement indicates that our prediction model successfully fuses the complementary and important metastasis information from separate feature categories. Thus, an accurate and reliable marker can be created to facilitate the diagnosis of LN metastasis. When investigating the contributions of separate feature groups in the final principal components, we observe that high performance feature groups (e.g., Shape and Density, Wavelet-DB2, Pyramid-FD, Wavelet-FD, WL-Haar) correspondingly have relatively large PCA coefficient weights. The performance of Wavelet-DB2 feature alone is only slightly lower than Shape and Density (0.737 vs 0.740), but Wavelet-DB2 exhibits significantly larger coefficients in the final components. This phenomenon shows that these two types of the features are somewhat homogeneous and Wavelet-DB2 features become dominant during the component synthesis. Another notable observation is the LBP feature has significantly higher coefficients as compared to its relatively low discriminative power, which may be attributed to the complementary information that are missed by the other high performance feature. In addition, LBP is used to identify and extract the local tumor texture, this again indicates the importance of local texture change in metastasis prediction.

Although the study results are encouraging, we also recognize that this investigation has several limitations as follows. First, the dataset only consists of 127 patients, which are selected from only one hospital with the limited diversity. We have not investigated whether the model optimized in this study can be directly applied on the CT images acquired from other medical centers, which will be obtained using different models of CT machines with the possibly different in signal-to-noise ratios (SNR). To further verify the performance and robustness of our propose model, a more comprehensive patient cohort should be established to include the patients from multiple medical centers. Second, we designed a very large feature pool (1,763 features) for the metastasis prediction in this study, but only one linear dimension reduction method (i.e. PCA) was used to identify the effective and non-redundant feature clusters. Some non-linear algorithms, such as sequential forward feature selection (SFFS) [41] and genetic optimization [42], should also be tested and compared with the PCA to further identify the meaningful information within the initial feature pool. Third, this preliminary study used the conventional prediction models for marker generation. The prediction performance may be further enhanced if we employ more advanced machine learning technologies. For example, the state of the art deep neural networks [43] can also be used as a fixed feature extractor to expand the existing feature pool, and the deep stacked auto-encoder (SAE) [44] can be applied for feature selection. As a result of the progress of applying deep learning in cancer image analysis [45, 46], we believe that the prediction accuracy may be further improved by combining the radiomics and deep learning based image feature together in future studies.

Fourth, we did not investigate the reproducibility of the computed tumor features, as they may vary due to the image quality or noise level changes when using different CT imaging machines or acquisition protocols [47]. For example, the changes oe adjustment of milliamperage (mAs level) changes radiation dose and impact image contrast-to-noise ratios. Increase of image noise may cause segmentation differences or errors on cervical tumors and have impact on computing some tumor-related features. In order to identify highly invariant or less sensitive to noise features, a phantom study [48] can be an approach to select the most robust features which are insensitive to the change of image acquisition parameters (i.e., mAs and KVP). Finally, although 2D image features are more robust with low generalization errors, 3D features may carry more prognostic information with higher discriminatory power. To validate this hypothesis, more research effort s are needed, which aim to: 1) develop accurate segmentation algorithms to reduce the errors in multiple slice tumor segmentation; 2) select the 3D features which are resistive to the change of scanning parameters (e.g., axial scanning resolution); and 3) minimize the variance of the scanning protocols at different clinical sites.

In summary, this study demonstrates that radiomics and machine learning provides powerful tools to develop new quantitative imaging markers. However, the scientific rigor of these new markers still needs to be further tested and validated using more independent and comprehensive image datasets in future studies.