Support Vector Machines Model of Computed Tomography for Assessing Lymph Node Metastasis in Esophageal Cancer with Neoadjuvant Chemotherapy.

OBJECTIVE: The aim of this study was to diagnose lymph node metastasis of esophageal cancer by support vector machines model based on computed tomography.
MATERIALS AND METHODS: A total of 131 esophageal cancer patients with preoperative chemotherapy and radical surgery were included. Various indicators (tumor thickness, tumor length, tumor CT value, total number of lymph nodes, and long axis and short axis sizes of largest lymph node) on CT images before and after neoadjuvant chemotherapy were recorded. A support vector machines model based on these CT indicators was built to predict lymph node metastasis.
RESULTS: Support vector machines model diagnosed lymph node metastasis better than preoperative short axis size of largest lymph node on CT. The area under the receiver operating characteristic curves were 0.887 and 0.705, respectively.
CONCLUSIONS: The support vector machine model of CT images can help diagnose lymph node metastasis in esophageal cancer with preoperative chemotherapy.

The prognosis of patients with resectable esophageal cancer remains poor. The reported 5-year survival rates range from 20 to 36% after intentionally curative surgery.[1-3] Median survival is only 9–24 months in patients with surgical treatment.[4-9] Prospective randomized trials demonstrated an improved survival after neoadjuvant therapy compared to surgery alone in patients with esophageal cancer. Some important studies include the CROSS trial, which analyzed neoadjuvant chemoradiation for patients with esophageal adenocarcinoma or squamous cell carcinoma, and the MAGIC and French trials analyzing neoadjuvant chemotherapy for adenocarcinoma.[10-12] Data from the FFCD9901 study suggested preoperative chemoradiotherapy increases complication incidence and mortality.[13] Therefore, preoperative chemotherapy in treating esophageal carcinoma is gradually accepted by surgeons.

As reported by Worldwide Esophageal Cancer Collaboration, survival decreases with the presence of lymph node metastases (LNM).[14] Indeed, the lymph node category was shown to be an independent prognostic factor in lymph node positive patients with resectable thoracic esophageal cancer.[15] Imaging examinations are the most commonly used tools for lymph node status evaluation in esophageal cancer. Wakelin et al compared computed tomography (CT), laparoscopic ultrasound, and endoscopic ultrasound (EUS) in the preoperative staging of esophagogastric carcinoma; the accuracy of CT in diagnosing the N stage of esophageal cancer was 17 of 29 (59%).[16] These authors concluded that the nodal status remains the most difficult area to assess using all three modalities. The main hurdle appeared to be the differentiation between benign and malignant enlarged nodes, with lymph node size alone not being a good criterion for assessing malignancy.

In recent years, machine-learning methods have been used to predict complex biological problems. Support vector machines (SVMs) are supervised machine learning techniques widely used in pattern recognition and classification problems. An SVM algorithm performs a classification by constructing a multidimensional hyperplane that optimally discriminates between two classes, by maximizing the margin between two data clusters. This algorithm achieves high discriminative power by using special nonlinear functions called kernels to transform the input space into a multidimensional space.[17] SVMs have been used in medical applications.[18-20] Given a set of training cases, each marked as belonging to one of two categories, a SVM training algorithm builds a model that predicts whether a new case falls into one category or the other.

Therefore, we used CT imaging data before and after neoadjuvant chemotherapy to establish a SVM mathematical model. In addition, the diagnostic power of the SVM method for differentiating LNM in patients with esophageal cancer was assessed. Because squamous cell carcinomas are significantly more common than adenocarcinomas and other malignant esophageal cancers in Asians, only patients with squamous cell carcinomas were evaluated in this study.

MATERIALS AND METHODS

Patients

This retrospective study was approved by the Ethics Committee of our hospital, with a waiver of requirement for informed consent. The clinical data were collected from the prospective database of our hospital. All patients in this database who had pathologically confirmed esophageal squamous cell carcinoma and received preoperative chemotherapy from January 2006 to January 2012 were included. All patients underwent gastroscopy to acquire pathological information, and received baseline and preoperative enhanced CT examinations.

Exclusion criteria were (a) pathologically proven adenocarcinoma, small cell carcinoma, mixed cancer, or other diseases; (b) other preoperative therapies (e.g., radiotherapy) simultaneously; (c) esophageal multiple primary carcinoma; (d) death within 30 days after surgery; (e) enhanced CT data before preoperative chemotherapy not obtained or images not interpretable; and (f) non-suitability for radical esophagectomy because of tumor progression or patient’s physical condition.

CT Protocol

MDCT was performed using a 64–detector row CT scanner (LightSpeed 64; GE Healthcare, Milwaukee, WI). Chest unenhanced CT scans were acquired with 0.625 mm collimation, 120–140 kVp, and 300–350 mAs. Subsequently, a total of 100 ml iopromide (Ultravist; Schering, Berlin, Germany) was administered intravenously via an 18-gauge angiographic catheter inserted into an antecubital vein, at 3 mL/sec with an automatic injector. Contrast-enhanced CT scans were performed at 60 seconds after iopromide injection. Sagittal and coronal reconstructions were carried out with contrast-enhanced images.

Image Analysis

Baseline and preoperative CT images were analyzed using the picture archiving communication system (PACS) by two independent radiologists blinded to patients’ clinical history. The following CT indicators were measured:

Tumor length: longest diameter obtained from the sagittal CT image.

Tumor thickness: lesion thickness obtained from the axial CT image.

Tumor CT value: region of Interest (ROI) placed on the lesion with maximum cross-section at the cross-sectional CT image.

Total LN number: number of all visible regional lymph nodes on Chest CT image.

Long axis size of largest regional LN (LSDL): long axis diameter of the largest regional lymph node in the axial CT image.

Short axis size of largest regional LN (SSDL). Diameter perpendicular to the long axis of the largest regional lymph node in the axial CT image.

The average results from the two radiologists were used for continuous variable analysis. Changes of CT image indicators between baseline and preoperative CT were assessed.

Statistical Analysis

LNM Assessment

All patients were divided into positive-LNM and negative-LNM groups, respectively. Node metastasis was confirmed by postoperative pathological results. A univariate statistical analysis with the SPSS software version 17.0 (SPSS Inc., Chicago, IL) was performed to evaluate the differences in various imaging indicators between the positive-LNM and negative-LNM groups. Group comparison was carried out by independent-samples T test. P <0.05 was considered statistically significant.

The CT indicators significantly different between positive-LNM and negative-LNM groups were selected to build the SVM model. Receiver operating characteristic (ROC) curves were used to evaluate these indicators in diagnosing LNM. The MedCalc software version 11.2 (MedCalc; MedCalc Software, Ghent, Belgium) was used to generate and compare the ROC curves.

Least Squares Support Vector Machine (SVM)

Least squares support vector machine (LS-SVM) was proposed by Suykens and Vandewalle.[21] Compared with other SVMs, LS-SVM utilizes quadratic sum of the slack variables as the penalty factor which ensures that LS-SVM can obtain a small training error. Specifically, LS-SVM minimizes the follow optimization problem, that is:

where x is the i-th training sample, y∈{−1,1} is the label of x, ϕ is a feature map which maps x to the feature space, w is the weight parameter vector, b is the bias parameter, e is i-th slack variable and e = (e1,⋯,e), y is a tuning parameter which makes a trade-off between the slack variable penalty and the margin. The Lagrangian function of Eq. (1) is

where α is i-th Lagrange multiplier (i = 1, ⋯, N). According to the optimal conditions, we have

and

By eliminating w and e, we can obtain the system of linear equations

where α = (α1,⋯,α), y = (y1,⋯,y), 1 = (1,⋯,1) is a length-N vector, I is N × N identity matrix, and Ω = (ϕ(x)ϕ(x)) is Gram matrix. Solving Eq. (7), we obtain the solution b* and α*. Then the optimal weight parameter vector w can be computed by Eq. (3) and e by Eq. (5). For a testing sample x, its label can be estimated by

LS-SVM can easily be extended to K-class (K > 2) classification problem by one-versus-one strategy.[22]

In this study, a free LS-SVM software package in MATLAB (version 7.0; MathWorks, Inc., Natick, MA) was applied to generate the SVM model. The default kernel function is used. A LS-SVM model for assessing LNM was established. Input indexes were the indicators obtained by the above univariate statistical analysis. The output index was lymph node metastasis in the patient. Confirmation was carried out by surgery and histopathology. Positive LNM was defined as 1 and negative LNM as −1. Fifty percent of cases were randomly selected to constitute the training sample. The remaining 50% of cases formed the testing sample. The training sample was used to establish the LS-SVM model. Finally, the ability of the model to predict LNM in the testing set and all cases was evaluated by ROC curves. ROC curves can be created automatically by the MATLAB software.

RESULTS

A total of 131 patients (102 males and 29 females; mean age of 58.0 years, ranging from 42 to 75 years) were included in the study (Fig. 1). There were 51 cases with lymph node metastasis and 80 without. The clinicopathological features of the patients are detailed in Table 1. The majority of patients (97%; 127/131) received a platinum-based two-drug combination, mainly paclitaxel (175 mg/m2, IV, d1 Q21) and cisplatin (25 mg/m2 IV, d1–3 Q21); the remaining patients received nedaplatin (80 mg/m2) combined with paclitaxel. A total of one to four chemotherapy cycles were administered before surgery at 3–6 weeks after neoadjuvant chemotherapy.

FIGURE 1

Flow chart of the study.

TABLE 1

Patient Characteristics

In the univariate analysis, preoperative tumor thickness, preoperative long axis and short axis sizes of largest lymph node, total numbers of lymph nodes in baseline and preoperative CT, and change of tumor thickness in second CT showed statistically significant differences between the LNM positive and negative groups (Table 2). Of these six CT indicators, preoperative short axis size of largest lymph node yielded the highest power for diagnosing LNM in ROC curves (Table 3, Fig. 2), with an area under the ROC curve (AUC) of 0.705.

TABLE 2

The Results of Univariate Statistical Analysis for CT Indicators of Baseline and Preoperative CT

TABLE 3

AUC of CT Indicators

FIGURE 2

Receiver operating characteristic (ROC) curve for lymph node metastasis with six CT indicators. The highest AUC of these six CT indicators was 0.705 which was performed by the short axis size of maximum lymph node (SSLN) of preoperative CT. Figure 2 can be viewed online in color at www.jcat.org.

After the random sampling by the SPSS statistical software, 66 cases were randomly selected to constitute the training sample. The other 65 cases were defined as testing sample. The training sample was used to establish the LS-SVM model. When we use this model to predict the training sample and testing sample, the AUCs of the SVM model were 0.955 and 0.553, respectively (Fig. 3a, b). The AUC of the SVM model predicting all samples reached 0.887 (Fig. 3c).

FIGURE 3

(A–C) Receiver operating characteristic (ROC) curve for lymph node metastasis with LS-SVM model. A, The AUC of the model predicting training sample was 0.955. B, The AUC of the model predicting testing sample was 0.553. C, The AUC of the SVM model predicting all samples reached 0.887. Figure 3 can be viewed online in color at www.jcat.org.

By comparing the ROC curves, the SVM model performed significantly better than preoperative short axis size of largest lymph node (P < 0.05).

DISCUSSION

Lymph node metastasis affects the surgical treatment of patients with esophageal cancer, and is also an important prognostic factor. Currently, preoperative diagnosis mainly depends on various imaging methods. Yokota et al indicated that clinical node diagnosis has low specificity and negative predictive value for predicting pathological nodes in the preoperative diagnosis of lymph node metastasis for patients with locally advanced resectable esophageal cancer.[23]

Because metastatic lymph node detection on CT images mainly depends on size, the sensitivity and specificity of CT vary with the definition of an abnormally enlarged node. In general, intrathoracic and abdominal lymph nodes greater than 1 cm in diameter are considered to be enlarged, while supraclavicular lymph nodes with a short axis exceeding 5 mm are considered to be pathologic.[24]

Most studies using the common size criterion of 1 cm to define enlarged nodes report CT sensitivity of 30–60%, whereas specificity tends to be somewhat higher (60–80%).[25,26] In a study by Picus et al, nearly all metastatic peri-esophageal lymph nodes measuring less than 7 mm were indistinguishable from non-metastatic lymph nodes by CT.[27] In addition, the presence of benign enlarged and inflammatory lymph nodes in esophageal cancer reduces the specificity of CT for detecting lymph node metastases. The lack of uniform criteria is the main constraint in predicting lymph node metastasis preoperatively.

The biological behavior of esophageal cancer reflects the histopathological performance of tumor malignancy and invasion. It affects lymph node metastasis directly or indirectly. The concrete manifestations of cancer biological behavior include, for example, tumor size, tumor invasion of other organs, lymph node metastasis, and distant metastasis. Therefore, MDCT imaging can accurately reflect the biological behavior of esophageal cancer histopathology. Univariate analysis in this study showed that all six indicators obtained from CT images were associated with LNM in esophageal cancer. Therefore, these biological behavior factors should be taken into account in comprehensively predicting LNM.

Other machine-learning methods have been used in medical studies. The mainly used method is artificial neural network (ANN), which is considered an appropriate tool for medical data analysis.[28] Bollschweiler et al applied a single-layer perceptron, an ANN, to predict lymph node metastasis in esophageal cancer, with an accuracy of 79%.[29] However, the ANN has some disadvantages: (1) the model is prone to overfitting, (2) it requires lengthy development and optimization time, and (3) it is more difficult to use in the field because of computational requirements.[30] Considering the above reasons, this study instead selected the SVM model, which could produce lower prediction error compared with classifiers based on other methods like artificial neural networks.[31] Compared with ANN, SVM may have the same or even better predictive ability.[32,33] Few reports are available regarding the application of SVM in esophageal cancer lymph node metastasis. In this preliminary study, the results indicated that the SVM model has better diagnostic capability for LNM than the traditional LN size criteria, with AUC achieving a good diagnostic power. With further improvement, SVM may become an effective tool in predicting lymph node staging in esophageal cancer.

Our study has some limitations. First, although a relatively large sample size was used, this was a single-center retrospective study. Further prospective studies are warranted to confirm the diagnostic power of the SVM model. In addition, the majority of patients were male (77.9%). Gender factors may influence the external validity of these findings. Finally, the AUC obtained for the LS-SVM model in the testing sample was relatively lower compared with the training sample value. This indicates a need for improvement of the model’s ability to assess new cases.

CONCLUSIONS

The least squares support vector machine model based on CT images can help diagnose lymph node metastasis in esophageal cancer with preoperative chemotherapy.