A prediction model for degree of differentiation for resectable locally advanced esophageal squamous cell carcinoma based on CT images using radiomics and machine-learning

Abstract

Objective:

To propose the prediction model for degree of differentiation for locally advanced esophageal cancer patients from the planning CT image by radiomics analysis with machine learning.

Methods:

Data of 104 patients with esophagus cancer, who underwent chemoradiotherapy followed by surgery at the Hiroshima University hospital from 2003 to 2016 were analyzed. The treatment outcomes of these tumors were known prior to the study. The data were split into 3 sets: 57/16 tumors for the training/validation and 31 tumors for model testing. The degree of differentiation of squamous cell carcinoma was classified into two groups. The first group (Group I) was a poorly differentiated (POR) patients. The second group (Group II) was well and moderately differentiated patients. The radiomics feature was extracted in the tumor and around the tumor regions. A total number of 3480 radiomics features per patient image were extracted from radiotherapy planning CT scan. Models were built with the least absolute shrinkage and selection operator (LASSO) logistic regression and applied to the set of candidate predictors. The radiomics features were used for the input data in the machine learning. To build predictive models with radiomics features, neural network classifiers was used. The precision, accuracy, sensitivity by generating confusion matrices, the area under the curve (AUC) of receiver operating characteristic curve were evaluated.

Results:

By the LASSO analysis of the training data, we found 13 radiomics features from CT images for the classification. The accuracy of the prediction model was highest for using only CT radiomics features. The accuracy, specificity, and sensitivity of the predictive model were 85.4%, 88.6%, 80.0%, and the AUC was 0.92.

Conclusion:

The proposed predictive model showed high accuracy for the classification of the degree of the differentiation of esophagus cancer. Because of the good prediction ability of the method, the method may contribute to reducing the pathological examination by biopsy and predicting the local control.

Advances in knowledge:

For esophageal cancer, the differentiation of degree is the import indexes reflecting the aggressiveness. The current study proposed the prediction model for the differentiation of degree with radiomics analysis.

Introduction

Esophageal cancer has been high mortality and incidence rate for many countries.¹ In Asia, the most common type of esophageal cancer is differentiated squamous cell carcinoma (SCC).² Neoadjuvant chemoradiotherapy (NCRT) followed by surgery has become the standard treatment modality for patients with resectable locally advanced esophageal cancer.³ The past clinical studies of NCRT for esophagus cancer reported that the local control predicts better disease control and survival outcomes.⁴ For NCRT, the predictive models of treatment response can provide crucial information before surgery. Liu et al reported the TNM stage and differentiation degree has been the import indexes reflecting the aggressiveness of the esophagus cancer.⁵

The important indexes reflecting the aggressiveness of esophageal cancer have been evaluated with endoscopic ultrasonography (EUS), and endoscopic biopsy preoperatively. The EUS was useful for the classification of the T staging, and endoscopic biopsy pre-operatively could evaluate differentiation degree.⁶ Levine et al investigated the safety experience of systematic endoscopic biopsy protocols for patients with Barrett’s esophagus.⁷ The rate of complication was 1.6% for all patients, which was higher than approximate 0.5% of complication rate for routine diagnostic endoscopy.⁸ Recently, the introduction of radiomics and artificial intelligence (AI) into medicine has become a major topic. Radiomics is the science that systematically handles a large amount of image information in radiology. By combining radiomics and AI, attempts have been made to automatically perform image diagnosis and predict the malignancy of cancer or treatment results.⁹ Thus, radiomics can be utilized to avoid undesirable complications caused by biopsy.

The purpose of this study is to propose the prediction model for the degree of differentiation for locally advanced esophageal cancer patients from the planning CT image by radiomics analysis with machine learning.

Methods and materials

The process of the radiomics analysis with machine learning is shown in

A) Patients

All patients who underwent the surgery after the NCRT from 2003 to 2016 at Hiroshima University hospital were used for the analysis; a histologically confirmed thoracic esophageal or esophagogastric junction (EGJ) cancer; age ≤75 years; a performance status of 0 to 2 according to the World Health Organization scale; a histologically confirmed thoracic esophageal or esophagogastric junction (EGJ) cancer. All patients provided written informed consent for treatment. Our institutional review board approved this retrospective study (E-1656–1). The characteristics of the patients and their tumors are presented in Table 1.

Table 1.

Patient characteristics

Gender	M/F	88/16
Age	Median (range)	66 (34–78)
PS	0/1/2	24/49/31
Site	Ut/Mt/Lt-Ae	24/49/27
Stage	IB/II/III/IV	6/24/59/15

B) Image acquisition

The workflow of the predicting model was shown in Figure 1. CT imaging of a patient was performed during free breathing on a CT scanner (Lightspeed RT16, GE Healthcare, Little Chalfont, UK). The slice thickness and slice interval were 2.5 mm. A tube voltage of 120 kVp and a current of 100–600 mA was used for imaging acquisition.

Figure 1.

The process of the radiomics analysis and create prediction model.

C) Contouring

Radiotherapy was performed with three-dimensional radiotherapy treatment planning for all patients. The primary tumor was defined as “tumor”, and the tumor with a 5 mm margin in all directions was defined as “tumor + 5 mm”. The shell feature was extracted from the voxels around the tumor boundaries, which was defined as the region between tumor + 5 mm and tumor the current study. The inner tumor feature was extracted from the voxels excluded the tumor boundaries. It was defined as the region of the tumor excluding outer 2 mm. All of the segmentations for radiomics analysis except for primary tumor were extended or reduced from the primary tumor. The primary tumor was segmented by one or two radiation oncologists who include one expert radiation oncologist at least. After the segmentation, the other expert radiation oncologist confirmed the segmentation. Moreover, two analyzers confirmed the segmentation before and after the deformable image registration.

F) Radiomics analysis

The extracted radiomics feature was standardized using z-score normalization so that each feature has the same mean of 0 and a standard deviation of 1 as follows:

$$f\left(x\right)=\frac{x-{\mu }_x}{{\sigma }_x}$$ (1)

where f(x) and x are the standardized and original voxel intensities, respectively, and μ_x is the mean standard deviation (SD) and σ_x is the SD of the image intensity values per case, respectively.

The entire data set of the CT was analyzed to extract a number of textural features from the segmentations for the radiotherapy plans. The feature extraction was performed using an open-source package in Python, Pyradiomics software.¹⁰ A detailed list of the radiomics features is shown in Supplementary Material 1.

A list of 50 quantitative features including first-order features: 21 features, shape features: 13 features, texture analysis features such as Gray Level Size Zone Matrix (GLSZM) features, and Features Gray Level Run Length Matrix (GLRLM) features: 93 features. were extracted. Moreover, we performed the radiomics analysis for the CT image with the imaging filters. These are preprocessed with a wavelet imaging filter. The wavelet filter has low-pass (L) and high-pass (H) filters. The decompositions are constructed in x, y, and z-direction. For example, HLL is then interpreted as the high-pass sub band, resulting from directional filtering of X with a high-pass filter along the x-direction, a low-pass filter along y-direction and a low-pass filter along the z-direction. In the current study, eight decompositions wavelet-HLL (wavelet-LHL, wavelet-LHH, wavelet-LLH, wavelet-HLH, wavelet-HHH, wavelet-HHL, wavelet-LLL) were performed. Each feature was computed with each of the above-mentioned preprocessing steps separately. From above, a total of 837 features were analyzed for each segmentation in this study.

G) Prediction model

To prevent overfitting, the least absolute shrinkage and selection operator (LASSO) regression model, which is suitable for the regression of high-dimensional data, was used with MATLAB code.^11,12 The LASSO performs feature selection during model construction by penalizing the respective regression coefficients. As this penalty is increased, more regression coefficients shrink to zero resulting in a more regularized model. The most significant predictive features were selected from among all the candidate features in the training set.

The objective of this study was to stratify patients into two labeled classes using different ML classifiers. In this regard, POR patients were labeled as 1, and well and moderately differentiated patients were labeled as 0. This issue also was repeated to classify patients with stages. Machine-learning based classification was performed with neural-network (NN) method. The NN method was built with 10 hidden layers and rectified linear unit activation (ReLU), as implemented in the MATLAB program. Moreover, the performance of the prediction with NN method was evaluated by comparing with conventionally used models: Classification and Regression Tree (CART), Support Vector Machine (SVM), and K-nearest Neighbor (KNN), as implemented in the MATLAB program.

The radiomics analysis can deal with the texture features with the wavelet filter in addition to the traditional semantic features such as the shape and histogram analysis. The usability of the prediction model using all radiomics features was evaluated by comparing with the prediction model using the traditional semantic features. The traditional semantic features used the shape and histogram methods, which were listed in Supplementary Material 1. The prediction model was built without the feature selection and the machine learning was used the NN method.

Here, all patients were randomly partitioned into a training/validation data set (57/16 patients) for generation of the model, and a test data set (31 patients) for the testing model performance. As shown Figure 2, the training–validation–testing processes were repeated five times for the fivefold cross-validation for all ML-methods. The predictive performance of all models was evaluated using the area under the curve (AUC) of receiver operator characteristic (ROC).

Figure 2.

Generation of the prediction model with the training/validation dataset and testing with the test data set. The fivefold cross-validation was applied for robust estimation.

Results

A total of 3480 features were extracted from CT image and were finally reduced to 11 features with the LASSO regression model, as shown in Figure 3. Table 2 shows the list of the selected features by LASSO regression. 21 features were selected with the wavelet filters and one feature was selected without an imaging filter. The features were extracted from various segmentations: three features from the tumor, two features from the tumor + 5 mm, three features from the shell region, two features from the inner region. nine features were selected from the texture analysis of GLCM, NGTDM, GLRLM, and GLDM. two features were selected from the first -order analysis.

Figure 3.

Radiomics feature selection using the Lasso logistic regression model. (a) Tuning penalization parameter λ (λ) and minimum criterion in the LASSO model. The binomial deviance was plotted vs log (λ). (b) Lasso coefficient profiles of the 3480 radiomics features. The green line showed the optimal λ in the LASSO method with the least partial likelihood deviance. LASSO, least absolute shrinkage and selection operator.

Table 2.

Selected the features by the LASSO regression

Region	Filter	Category	Item
Tumor	wevelet-LHL	NGTDM	Contrast
Tumor	wavelet-HLH	GLDM	Large Dependence Low Gray Level Emphasis
Tumor	wavelet-HHL	Firstorder	Mean
Tumor +5 mm	wavelet-HLL	GLRLM	Run Variance
Tumor +5 mm	original	GLCM	Idn
Inner	wavelet-HLH	GLDM	Large Dependence Low Gray Level Emphasis
Inner	wavelet-HLH	GLCM	MaximumProbability
Shell	wavelet-HLL	GLCM	Joint Average
Shell	Wavelet-LHL	Firstorder	Mean
Shell	Wavelet-LHH	GLCM	Idn
Shell	Wavelet-HLH	GLDM	Small Dependence Emphasis

LASSO, least absolute shrinkage and selection operator.

Those models were evaluated with the training and test data sets with fivefold cross-validation, as shown in Table 3. Figure 4 shows the performance of the predictive model was tested according to the ROC metrics for the test dataset on fivefold cross-validation. The average accuracy of the five models was 87.5%, with the training data set. The average accuracy, sensitivity, and specificity with the test data were 85.4%, 88.6%, and 80.0%, respectively. Figure 4 shows the performance of the classifier which was validated according to the ROC metrics in the test data set with fivefold cross-validation. The AUC was 0.90 for first model, 0.92 for second model, 0.93 for third model, 0.91 for fourth mode, and 0.92 for fifth model. The average and standard deviation AUC with fivefold cross-validation was 0.92 and 0.0.

Table 3.

Model performance of NN with fivefold cross-validation in the accuracy, sensitivity, and specificity for the training and test data set

		Model I	Model II	Model III	Model IV	Model V	Average	SD
Training	Accuracy (%)	89.1	94.5	82.8	83.6	87.3	87.5	3.8
Test	Accuracy (%)	88.0	82.6	87.0	82.6	87.0	85.4	2.1
	Specificity (%)	89.5	84.6	92.9	82.4	93.8	88.6	4.1
	Sensitivity (%)	87.5	80.0	77.8	83.3	71.4	80.0	4.9
	AUC	0.90	0.92	0.93	0.91	0.92	0.92	0.0

AUC, area under the curve; SD, standard deviation.

Figure 4.

The performance of the predictive model with NN was tested according to the ROC metrics for the test dataset on fivefold cross-validation. ROC, receiver operator characteristic.

Figure 5 shows average ROC metrics with the performance of the predictive model of CART, SVM, and KNN for the test data set. Table 4 shows the average accuracy, sensitivity, specificity, and AUC for the model performance of NN, CART, SVM, and KNN for the test dataset. The accuracy, sensitivity, specificity, and AUC were significantly highest for the model of NN.

Figure 5.

Average ROC metrics with the performance of the predictive model of CART, SVM, and KNN for the test data set. CART, Classification and Regression Tree; KNN, K-nearest Neighbor; ROC, receiver operator characteristic; SVM, Support Vector Machine.

Table 4.

Model performance of the NN, CART, SVM, and KNN with fivefold cross-validation in the accuracy, sensitivity, specificity, and AUC for the test data set

	NN	CART	SVM	KNN
Accuracy (%)	85.4	76.1	78.2	77.1
Specificity (%)	88.6	68.9	74.1	82.7
Sensitivity (%)	80.0	79.3	80.0	67.6
AUC	0.92	0.69	0.85	0.76

AUC, area under the curve; CART, Classification and Regression Tree; KNN, K-nearest Neighbor; NN, neural network; SVM, Support Vector Machine.

Figure 6 shows average ROC metrics with the performance of the predictive model using all radiomics features and traditional semantic features for the test data set. Table 5 shows the comparison of the prediction performance of NN using all radiomics features and traditional semantic features with fivefold cross-validation for the test data set. The performance of the prediction model using all radiomics features was superior to that using the traditional semantic features.

Figure 6.

Average ROC metrics for the performance of the predictive model of NN with the traditional semantic features for the test data set. NN, neural network; ROC, receiver operator characteristic.

Table 5.

Model performance of the NN with all radiomics analysis features and traditional semantic features with fivefold cross-validation in the accuracy, sensitivity, specificity, and AUC for the test data set

	All radiomics features	Traditional semantic features
Accuracy (%)	85.4	55.7
Specificity (%)	88.6	61.8
Sensitivity (%)	80.0	36.0
AUC	0.92	0.50

AUC, area under the curve; NN, neural network.

Discussion

Liu et al investigated the correlation of the texture parameters and T, N staging, and differentiation degrees with CT image.⁵ The parameters of kurtosis and entropy were significantly correlated with T stage, and some parameters (kurtosis, skew, mean, 10th and 90th percentile) are correlated to N stage. On the other hand, the correlation of the texture parameters and differentiation degree was weak.

The traditional semantic features can extract shape, physical characteristics, and statistical information. The accuracy of the prediction performance was significantly improved by adding the texture features. Radiomics analysis extracts a large number of quantitative image features, which leads to comprehensive quantification of tumor phenotypes.¹³ In the current study, the predictive model demonstrated good predictive power in both training (accuracy: 85.4%) and test (accuracy: 88.6%, AUC: 0.92) datasets. The current study proposed the predictive model for the differentiation degree with the CT-based radiomics feature in the shell region which is the outer of the tumor region and the inner tumor region which eliminated the peripheral of the tumor in addition to the tumor region. Hao et al proposed a predictive model for a distant failure by shell analysis that allows us to detect its associations with metastasis within the microenvironment.¹⁴ Usually, the segmentation of the tumor in treatment planning for radiotherapy was defined by a physician with visual assessment. The current study showed radiomics feature within the microenvironment could also be useful for predicting the differentiation degrees. Moreover, the current study used the imaging filter before extracting the radiomics features. For the feature selection with LASSO regression, the radiomics features with the wavelet filter and that extracted in texture analysis were mostly selected. Nie et al performed radiomics analysis in the liver for CT image to differentiating focal nodular hyperplasia (FNH) from hepatocellular carcinoma.¹⁵ 10 features which were high-order filter and wavelet features were selected as the significant predictors to construct the radiomics signature. The imaging filter could reduce noise, enhance the edge, and extract or eliminate a constant frequency. It leads to eliminate the redundant and extract effective factors for prediction.

The prediction performance with the NN-based machine learning was higher than the other models of the CART, SVM, and KNN. The NN-based machine learning uses the weights expressing the importance of the respective inputs which is the radiomics features to the output which is the group of the differentiation degree. Thus, the NN-based machine learning can combine the radiomics features.

Because of the good prediction ability of the method, the method could contribute to reducing the pathological examination by biopsy. Moreover, the predictive model for the differentiation of degree of the locally advanced esophagus cancer may use for the prediction of the local control without examination of bioscopy. There were two limitations in the current study. One was that the current study was performed in the single institution. The other was low sample size. Future investigations will use the data of large sample size collected from other institutions and validate the prediction model in a multi institutional study.

Conclusion

The proposed predictive model showed high accuracy for the classification of the degree of the differentiation of esophagus cancer. Because of the good prediction ability of the method, the method may contribute to reducing the pathological examination by biopsy and predicting the local control.