Radiomic analysis in T2W and SPAIR T2W MRI: predict treatment response to chemoradiotherapy in esophageal squamous cell carcinoma

Abstract

Background

To investigate the capability of radiomic analysis using T2-weighted (T2W) and spectral attenuated inversion-recovery T2-weighted (SPAIR T2W) magnetic resonance imaging (MRI) for predicting the therapeutic response of esophageal squamous cell carcinoma (ESCC) to chemoradiotherapy (CRT).

Methods

Pretreatment T2W- and SPAIR T2W-MRI of 68 ESCC patients (37 responders, 31 nonresponders) were analyzed. A number of 138 radiomic features were extracted from each image sequence respectively. Kruskal-Wallis test were performed to evaluate the capability of each feature on treatment response classification. Sensitivity and specificity for each of the studied features were derived using receiver operating characteristic (ROC) analysis. Support vector machine (SVM) and artificial neural network (ANN) models were constructed based on the training set (23 responders, 20 nonresponders) for the prediction of treatment response, and then the testing set (14 responders, 11 nonresponders) validated the reliability of the models. Comparison between the performances of the models was performed by using McNemar’s test.

Results

Radiomic analysis showed significance in the prediction of treatment response. The analyses showed that complete responses (CRs) versus stable diseases (SDs), partial responses (PRs) versus SDs, and responders (CRs and PRs) versus nonresponders (SDs) could be differentiated by 26, 17, and 33 features (T2W: 11/11/15, SPAIR T2W: 15/6/18), respectively. The prediction models (ANN and SVM) based on features extracted from SPAIR T2W sequence (SVM: 0.929, ANN: 0.883) showed higher accuracy than those derived from T2W (SVM: 0.893, ANN: 0.861). No statistical difference was observed in the performance of the two classifiers (P=0.999).

Conclusions

Radiomic analysis based on pretreatment T2W- and SPAIR T2W-MRI can be served as imaging biomarkers to predict treatment response to CRT in ESCC patients.

Introduction

Esophageal carcinoma (EC) is the 8th most commonly occurring types of malignancy, including more than 450,000 new cancer diagnoses yearly, and also the 6th leading cause of cancer-related mortality with an estimated approximately 400,000 deaths every year (1). Esophagostomy is an effective therapeutically method in the treatment of patients with early-stage disease. Unfortunately, most patients diagnosed with esophageal cancer present at locally advanced stage with unresectable or metastatic disease, for which concurrent chemoradiotherapy (CRT) is currently considered as the best palliative option, as it can improve survival and reduce the local recurrence rate in patients who respond to the treatment (2). On the other hand, patients who are insensitive to CRT (18–25%) may be harmed unnecessarily by the toxicity of an ineffective therapy without prognostic benefit (3). Therefore, early identification of patients who are at higher risk of poor response before CRT would allow personalization of their treatment.

Recent advancements in magnetic resonance imaging (MRI) protocols and techniques have resulted in better imaging quality with the potential to be applied in staging, radiation treatment planning, and the assessment of therapeutic response (4). Previous studies have shown the value of MRI in predicting treatment response for glioma (5), rectal cancer (6), and nasopharyngeal carcinoma (7). For EC, diffusion-weighted imaging (DWI) and dynamic contrast-enhanced MRI (DCE-MRI) can provide complementary information for the prediction of treatment response early to CRT, besides measurements of dimensional changes in tumor diameter or volume on anatomical MRI (8,9). For example, Aoyagi et al. showed that the apparent diffusion coefficient (ADC) value could be used as a biomarker to predict treatment response for patients with esophageal squamous cell carcinoma (ESCC) (8) and more recently, Heethuis et al. found that changes in tumor area-under-the-concentration time curve (tAUC) value derived from DEC-MRI are promising for prediction of histopathologic response to neoadjuvant CRT (nCRT) for EC (9). However, although the effect of these clinical indicators was demonstrated over the past years, little attention has been paid to the predictive capability of pretreatment anatomical MRI (T2W and SPAIR T2W). Therefore, new tools based on the aforementioned sequences are expected for the prediction of treatment response early before CRT.

Radiomics is a novel image analysis technology that has received much attention in recent years due to its innovative properties, which provide valuable information regarding to tumor phenotype (10-14). This innovative method focuses on the improvement of image analysis by converting medical images into high-dimensional quantitative features via post-processing techniques. It then involves comprehensive analyses of these features with clinical outcomes using reproducible and robust methodology. Recent studies have shown that radiomic biomarkers based on positron emission tomography (PET) and computed tomography (CT) are associated with staging, treatment response, and survival across a range of cancer type (15-20). However, compared to PET and CT, MRI provides excellent contrast between the tumor and its surrounding structures, which allows tumor margins to be contoured more accurately (4). To our knowledge, no published study has determined whether the treatment response to CRT in ESCCs could be early predicted by a radiomics method using anatomical MRI (T2W and SPAIR T2W).

Thus, the purpose of this study was to use radiomic features derived from T2W- and SPAIR T2W-MRI combined with supervised machine-learning algorithms to predict treatment response to CRT in ESCC patients.

Methods

Patient database

The retrospective database contained pretreatment MRI images from 68 patients who histologically diagnosed with ESCC at our institution from February 2015 to December 2016. This study was approved by the Nanjing Drum Tower Hospital’s ethics committee (No. 2015-063-09) and informed consent from the patients was provided. Two different MRI series were acquired from the same image location, including T2W- and SPAIR T2W-MRI. All the patients were considered to be inoperable and not received chemotherapy or radiotherapy before MRI scan. The other inclusion criteria were met: normal cardiac, pulmonary, and hematologic function. Endoscopic ultrasonography (EUS), chest/abdominal CT scan with contrast-enhanced and chest MRI scan were used for the staging of the patients according to the AJCC 7th. Baseline patient characteristics were summarized in Tables 1,2.

Table 1. Baseline characteristic of patients in training set.

Characteristics	Responders (n=23)	Nonresponders (n=20)	P value
Age, median [range] (years)	61 [50–76]	67 [41–82]	0.071*
Gender
Male/female	13/10	13/7	0.571**
TNM staging
T1/T2/T3/T4	1/3/16/3	1/2/14/3	0.989**
N0/N1/N2	8/10/5	7/8/5	0.961**
M0/M1	21/2	19/1	0.635**

*, independent-samples t-test; **, chi-square test.

Table 2. Baseline characteristic of patients in testing set.

Characteristic	Responders (n=14)	Nonresponders (n=11)	P value
Age, median [range] (years)	64 [53–74]	63 [54–76]	0.873*
Gender
Male/female	10/4	9/2	0.546**
TNM staging
T1/T2/T3/T4	2/4/7/1	1/3/6/1	0.978**
N0/N1/N2	4/7/3	3/4/4	0.686**
M0/M1	12/2	10/1	0.692**

*, independent-samples t-test; **, chi-square test.

CRT

All patients were irradiated with 2 Gy per fraction in 30 fractions for intensity modulated radiation therapy (IMRT). The dose prescriptions were designed to cover at least 96% of planning target volume (PTV). During the whole course of radiotherapy, patients underwent 2–3 cycles of synchronous chemotherapy (nedaplatin + docetaxel/paclitaxel).

Treatment evaluation

One month after the completion of the treatment, therapeutic response was assessed by using CT/MRI image with contrast, according to the Response Evaluation Criteria in Solid Tumors (RECIST) (21). Patients with complete response (CR) or partial response (PR) were considered responders, while patients with stable disease (SD) or progressive disease (PD) were classified as nonresponders

MR image acquisition and tumor segmentation

All patients underwent a pretreatment 3.0 T MR scan from the same scanner (Philips Achieva 3.0T X-series, Philips Healthcare, Best, The Netherlands) according to a standard clinical acquisition protocol as follows: T2W MRI [repetition time (TR)/echo time (TE): 554.4/80 millisecond (msec); flip angel, 90 degrees; matrix size, 512×512; slice thickness, 5 mm; in-plane resolution, 0.70 mm × 0.70 mm] and SPAIR T2W MRI (TR/TE: 569.2/70 msec; flip angel, 90 degrees; matrix size, 432×432; slice thickness 4 mm, in-plane resolution 0.87 mm × 0.87 mm). The regions of interest (ROIs) were semi-automatically delineated by two expert radiologists (S Li for ROI-1 and W Ren for ROI-2) for each sequence using commercially available software (MIM Software, Cleveland, Ohio, USA) and then reviewed by an experienced radiologist (J Yan). The contours for each ROI were drawn around the largest cross-sectional area of the tumor lesion avoiding fat, air, and bone (Figure 1).

Figure 1.

Region of interest (ROI) was contoured by two radiologists based on different MRI series, and corresponding 2D ROIs. (A) ROI-1 based on T2-weighted MRI; (B) ROI-2 based on T2-weighted MRI; (C) ROI-1 based on SPAIR T2-weighted MRI; (D) ROI-2 based on SPAIR T2-weighted MRI.

Image preprocessing

Although patients underwent MR examination using the same scanner, changes of parameter settings and voltage may result in different MR imaging intensity ranges (22,23). In order to correct for these variations, the contoured ROIs were normalized using a two-step process as requires: (I) image filtering and (II) gray-level quantization. The first step (image filtering) was performed using a Winner filter to reduce the image noise. The second step normalized the voxel values within the ROIs to a finite gray-level range as follows:

$${\text{V}}_{\left(\text{X}\right)}\text{=}\left[{\text{2}}^{\text{K}}\frac{{\text{I}}_{\left(\text{X}\right)}-{\text{ min}}_{\text{i}\in \text{Ω}}\text{i}}{{\max }_{\text{i}\in \text{Ω}}\text{i}-{\text{min}}_{\text{i}\in \text{Ω}}\text{i}}\right]$$

where 2^k represent the four different discrete values [16, 32, 64, 128], I is the intensity of the raw image, and Ω is the pixels set in the delineated volume. These steps reduced the noise in the image and normalized the intensity range across patients, which in turn facilitate the comparison of the radiomic features. Previous studies showed that no statistically significant difference was found in choosing different resampling values (22,24,25), and 64 discrete values were chosen in our study.

Radiomic feature extraction

Imaging Biomarker Explorer Software V1.0β (IBEX) (26) and MATLAB 2015a (Mathworks, Natick, MA, USA) were used to extract radiomic features from delineated two-dimensional (2D) ROIs. IBEX is an open software platform that aimed to support radiomics workflow tasks such as multimodality image data import/export, review, preprocessing, and feature extraction. Using IBEX software, several categories features were extracted, including intensity histogram (IH), gradient-orientation histogram (GH), gray-level co-occurrence matrix (GLCM; directions: 0°, 45°, 90°, 135° and mean; offset: 1 pixel distance), gray-level run-length matrix (GLRLM; directions: 0°, 90° and mean), and neighborhood intensity difference matrix (NIDM; neighborhood size: 3×3). In addition, our study contained more features than IBEX by adding Gabor wavelet transform (GWTF, five wavelet scales and eight filter orientations) (23), which performed using MATLAB software. All patients’ DICOMs files (MR images + ROI structures) were first exported from MIM software and then imported into the above modules to compute radiomic features.

In all, 138 radiomic features were extracted from each image sequence based on three principal methods: histogram-based (IH, GH), texture-based (GLCM, GLRLM, and NIDM), and transform-based (GWTF). More detailed contents of the extracted features are summarized in Table S1.

Inter-observer variability evaluation and feature reduction

Inter-observer variability of radiomic features extraction was assessed using intra-class correlation coefficient (ICC). It was performed to quantify the feature reproducibility in repeat delineation (ICC <0.40, poor agreement; 0.40≤ ICC <0.60, moderate agreement; 0.60≤ ICC <0.80, good agreement; ICC ≥0.80, excellent agreement). ICC analysis was performed using “irr” package (version 0.84) in R software (version 3.3.2) (27). In our study, radiomic features with ICC greater than 0.8 were extracted.

Additionally, pair-wise correlations of the studied features were also considered. For these highly reproducible features (ICC ≥0.80), a correlation matrix was created and absolute correlation coefficients (ACC) were computed using “caret” package (version 6.0-77) in R software. The ACC values close to 1 mean the features are correlated parameters. In our study, radiomic features with ACC greater than 0.8 were removed.

Statistical analysis

All statistical analyses were performed using R software version 3.3.2. The capability of each influential feature (ICC ≥0.8 and ACC <0.8) to classify patients with different treatment response were evaluated using the Kruskal-Wallis test (P<0.05), and Dunn-Bonferroni test was used for post hoc comparisons (24). Receiver operating characteristic (ROC) curves and area under the curve (AUC) analysis were used to assess the diagnostic efficacy of each studied features for discrimination among various treatment responses [95% confidence intervals (CIs), specificity and sensitivity were also calculated]. AUC with a value of 1 means an ideal result, while values lower than 0.5 indicates insignificant.

Feature selection and model construction

Radiomic features extracted from the two sequences (T2W and SPAIR T2W) were separately modeled using support vector machine (SVM) and artificial neural network (ANN) algorithms. All patients were randomized into two groups: 43 for training (23 responders, 20 nonresponders) and 25 for validating (14 responders, 11 nonresponders).

To reduce the training time and avoid model over-fitting, the feature dimension should be reduced firstly. Based on the training group, we chose wrapper-based feature selection method (28) to generate optimal feature subset for the specific model (SVM or ANN). This method separately performed on each feature group (features extracted from T2W and SPAIR T2W with ICC ≥0.8 and ACC <0.8) by recursively removing features and then assessing the predictive power of the remaining features without missing any critical ones. After processing by the above method, feature dimension was further reduced.

To assess classification performance, 10-fold cross-validation (CV) approach was served as the internal validation in the training group. The specificity, precision and accuracy of the predictions were derived from true positives (TP), false positives (FP), true negatives (TN), and false negatives (FN). Additionally, Matthews correlation coefficient (MCC) were also obtained to measure the predictive ability of the classifier. The MCC parameter ranged from −1 to +1, and values of +1 represented ideal prediction, 0 implied the equivalent of a random guess, and −1 indicated an inverse prediction.

Statistical comparison between ANN and SVM models

The statistical difference between the performance of different models was assessed using McNemar’s test (29). The test was performed on the outcomes obtained from the 10-fold CV.

Validation

Patients (n=25, 14 responders and 11 nonresponders) who were not involved in model construction were then served as an independent validation set. The confusion matrix contained prediction result was used to calculate the specificity, precision, and accuracy.

Results

Treatment response after CRT

Patients were evaluated 1 month after the completion of CRT. Thirty-one patients were classified as responders (31 SD, 0 PD), while 37 patients were considered as nonresponders (17 CR, 20 PR).

Predictive capabilities

As discussed above, 138 radiomic features were extracted from each image sequence. ICC and ACC were computed for all radiomic features. Based on the above-mentioned criteria (ICC ≥0.8 and ACC <0.8), 47 features were generated for the T2W and 39 features for the SPAIR T2W. Details of the features are summarized in Table 3.

Table 3. The features obtained from preprocessing.

Feature type	T2W	SPAIR-T2W
Histogram-based	IH_IQR, IH_Kurtosis, IH_Skewness, GH_MAD_2, GH_Skewness	IH_Kurtosis, IH_MAD_1, IH_Skewness, GH_IQR, GH_Skewness
Texture-based	GLCM_CS90°	GLCM_CS135°
	GLCM_Contrast0°/90°	GLCM_Contrast45°/90°/135°
	GLCM_Correlation0°/45°/90°/135°	GLCM_Correlationmean/0°/45°/135°
	GLCM_IDM90°	GLCM_IDM mean/0°/45°/90°
	GLCM_IV135°	GLCM_IV0°/45°/90°/135°
	GLRLM_GLN90°	GLRLM_LRE0°/90°
	GLRLM_LRE0°/90°	GLRLM_LRHGE90°
	GLRLM_LRHGE0°	GLRLM_LRLGE0°
	GLRLM_LRLGE0°	NIDM_Busyness
	GLRLM_SRLGE90°	NIDM_Coarseness
	NIDM_Busyness	NIDM_Complexity
	NIDM_Coarseness	NIDM_Contrast
	NIDM_Complexity	NIDM_TS
	NIDM_Contrast
	NIDM_TS
Transform-based	Gabor_MSA-14, -18, -21, -22, -26, -44, -51, -55	Gabor_MSE-11, -14, -18, -21, -22, -25, -26, -33, -57
	Gabor_MSE-11, -13, -14, -18, -21, -22, -26, -33, -38, -45, -51, -52, -56, -58

Result of Kruskal-Wallis test showed that 26 features (T2W: 11, SPAIR T2W: 15) could differentiate between CR and SD, 17 features (T2W: 11, SPAIR T2W: 6) could differentiate between PR and SD, and 33 features (T2W: 15, SPAIR T2W: 18) could differentiate between responders (CRs and PRs) and nonresponders (SDs). A summary of detailed results is shown in Tables 4,5. To differentiate responders from nonresponders, we analyzed the significant features with ROC curves and obtained the AUC values (T2W: 0.640–0.757, SPAIR T2W: 0.640–0.721). For example, GLCM_Contrast_0° (T2W-based), with an AUC of 0.677 (sensitive =0.903, specificity =0.459), could differentiate between responders and nonresponders, with a cutoff of 11.595 (Table 6), indicating that tumor lesions whose GLCM_Contrast_0° was higher than 11.595 were most likely from nonresponders. Similar results were obtained from the ROC analysis of SPAIR T2W-based features (Table 7).

Table 4. Features extracted from T2W-MRI that classify different treatment responses.

Feature type	Responders (CR, PR) versus nonresponders (SD)	CR versus SD	PR versus SD
Histogram-based	IH_IQR	IH_IQR	IH_IQR
Texture-based	GLCM_Contrast0°	GLCM_Contrast0°	GLCM_Contrast0°
	GLRLM_GLN90°	GLCM_Energy135°	GLCM_Energy135°
	GLRLM_LRE0°/90°	GLCM_IV135°	GLCM_IV135°
	GLRLM_LRHGE0°	GLRLM_LRE0°/90°	GLRLM_LRE0°/90°
	GLRLM_LRLGE0°	NIDM_Coarseness	GLRLM_LRHGE0°
	GLRLM_SRLGE90°	NIDM_Contrast	GLRLM_LRLGE0°
	NIDM_Coarseness	NIDM_Complexity	GLRLM_SRLGE90°
	NIDM_Contrast		NIDM_Coarseness
			NIDM_Contrast
Transform-based	Gabor_MSA-44; Gabor_MSE-45, -51, -52, -58	Gabor_MSA-14, Gabor_MSE-51	None

CR, complete response; PR, partial response; SD, stable disease.

Table 5. Features extracted from SPAIR T2W-MRI that classify different treatment responses.

Feature type	Responders (CR, PR) versus nonresponders (SD)	CR versus SD	PR versus SD
Histogram-based	IH_MAD_2	IH_MAD_2	None
Texture-based	GLCM_IDM0°;	GLCM_Contrast45°/135°	None
	GLCM_IV0°/45°/90°	GLCM_IDMmean/0°
	GLRLM_LRE0°/90°	GLCM_IV0°/45°/90°/135°
	NIDM_Coarseness	GH_Skewness
	NIDM_Contrast	GLRLM_LRE0°/90°
	NIDM_Complexity	NIDM_Coarseness
		NIDM_Contrast
		NIDM_Complexity
Transform-based	Gabor_MSE-11, -14, -18, -21, -25, -26, -33, -57	None	Gabor_MSE-18, -21, -22, -25, -26, -33

CR, complete response; PR, partial response; SD, stable disease.

Table 6. Features extracted from T2W-MRI show statistical difference between nonresponders and responders.

Feature	P value	Standard error	95% CI	AUC	Sens	Spec	Cut-off
GLCM_Contrast0°	0.012	0.065	0.552–0.785	0.677	0.903	0.459	>11.595
GLRLM_GLN90°	0.038	0.070	0.522–0.759	0.647	0.548	0.756	≤0.028
GLRLM_LRE0°/90°	0.022	0.066	0.537–0.772	0.662	0.709	0.594	≤1.656
	0.048	0.069	0.514–0.753	0.640	0.548	0.756	≤1.422
GLRLM_LRHGE0°	0.003	0.065	0.586–0.813	0.709	0.645	0.810	>9.037×103
GLRLM_LRLGE0°	0.001	0.062	0.601–0.824	0.723	0.741	0.648	≤4.426×10−4
GLRLM_SRLGE90°	<0.001	0.058	0.638–0.853	0.757	0.838	0.594	≤2.570×10−4
NIDM_Coarseness	0.021	0.067	0.538–0.773	0.663	0.645	0.675	≤0.634
NIDM_Contrast	0.028	0.067	0.530–0.766	0.655	0.548	0.783	>0.081
Gabor_MSA-44	0.013	0.068	0.551–0.784	0.676	0.741	0.621	≤0.058
Gabor_MSE-45	0.001	0.062	0.598–0.822	0.720	0.548	0.837	≤1.810×105
Gabor_MSE-51	0.042	0.069	0.518–0.756	0.643	0.516	0.810	≤5.316×104
Gabor_MSE-52	0.006	0.066	0.570–0.800	0.694	0.451	0.918	≤1.062×105
Gabor_MSE-58	0.002	0.064	0.590–0.816	0.712	0.838	0.594	≤3.393×105

CI, confidence interval; AUC, area under the curve; Sens, sensitivity; Spec, specificity; Responders, patients with CR and PR; Nonresponders, patients with SD; CR, complete response; PR, partial response; SD, stable disease.

Table 7. Features extracted from SPAIR T2W-MRI show statistical difference between responders and nonresponders.

Feature	P value	Standard error	95% CI	AUC	Sens	Spec	Cut-off
IH_MAD_2	0.009	0.063	0.559–0.790	0.683	0.677	0.567	>7
GLCM_IDM0°	0.030	0.066	0.528–0.764	0.653	0.838	0.486	≤0.996
GLCM_IV0°/45°/135°	0.019	0.066	0.540–0.775	0.665	1.000	0.378	≤0.241
	0.047	0.068	0.515–0.753	0.640	0.871	0.486	≤0.168
	0.034	0.066	0.524–0.761	0.650	0.967	0.324	≤0.193
GLRLM_LRE0°/90°	0.004	0.063	0.576–0.805	0.700	0.871	0.486	≤1.307
	0.003	0.062	0.587–0.813	0.710	0.741	0.621	≤1.264
NIDM_Coarseness	0.005	0.067	0.575–0.804	0.698	0.774	0.675	≤0.363
NIDM_Contrast	0.026	0.064	0.573–0.802	0.697	0.580	0.783	>0.178
NIDM_Complexity	0.005	0.066	0.532–0.768	0.657	0.580	0.702	>6.582×103
Gabor_MSE-11	0.010	0.067	0.558–0.790	0.682	0.387	0.945	>1.99×104
Gabor_MSE-14	0.039	0.073	0.521–0.758	0.646	0.483	0.891	>1.72×105
Gabor_MSE-18	0.046	0.069	0.515–0.754	0.641	0.387	0.891	>1.06×105
Gabor_MSE-21	0.001	0.062	0.599–0.823	0.721	0.741	0.648	>1.62×105
Gabor_MSE-25	0.003	0.064	0.584–0.811	0.707	0.838	0.621	>8.75×104
Gabor_MSE-26	0.002	0.062	0.595–0.820	0.718	0.935	0.459	>1.61×105
Gabor_MSE-33	0.014	0.065	0.549–0.782	0.673	0.580	0.729	>3.18×105
Gabor_MSE-57	0.048	0.069	0.514–0.753	0.640	0.387	0.891	>1.97×105

CI, confidence interval; AUC, area under the curve; Sens, sensitivity; Spec, specificity; Responders, patients with CR and PR; Nonresponders, patients with SD; CR, complete response; PR, partial response; SD, stable disease.

Supervised classification and statistical comparison

Wrapper-based feature selection method was separately performed on T2W- and SPAIR T2W-based feature sets (features with ICC ≥0.8 and ACC<0.8, Table 3) to obtain the optimal feature subsets for each model (SVM and ANN), with the results showing that 4 T2W-based/7 SPAIR T2W-based features were selected for SVM and 4 T2W-based/4 SPAIR T2W-based features were selected for ANN. As for T2W sequence, two features (GLRLM_LRE_0° and NIDM_Busyness) were both selected for these two models, and no histogram-based feature was selected. In the case of SPAIR T2W sequence, three features (GLCM_IV_0°, GLCM_Correlation_45° and Gabor_MSE-25) were both selected. Details of the optimal feature sets for each model are summarized in Table 8.

Table 8. Optimal feature set obtained from Wrapper-based feature selection.

Feature	T2W			SPAIR-T2W
	SVM	ANN		SVM	ANN
Histogram-based	None	None		GH_IQR	None
Texture-based	GLRLM_LRE0°	GLRLM_LRE0°		GLCM_IV0°	GLCM_IV0°/90°
	NIDM_Coarseness	NIDM_Busyness		GLCM_IDM0°	GLCM_Correlation45°
	NIDM_Busyness			GLCM_Correlation45°
				GLRLM_LRE90°
Transform-based	Gabor_MSA-44	Gabor_MSE-44, -52		Gabor_MSE-25, -33	Gabor_MSE-25

SVM, support vector machine; ANN, artificial neural network.

Table 9 summarizes the classification results obtained from training group by 10-fold CV, in terms of specificity, precision, weighted accuracy, and MCC for each model. Radiomic features derived from SPAIR T2W sequence (SVM: 0.929, ANN: 0.883) generally achieved higher predictive accuracy than those derived from T2W sequence (SVM: 0.893, ANN: 0.861).

Table 9. Summary of classification results obtained from training group by 10-fold CV on T2W and SPAIR T2W radiomic features.

Feature set	Algorithm	TP rate	FP rate	Specificity	Precision	Accuracy	F-measure	MCC
T2W	SVM	0.884	0.127	0.846	0.890	0.893	0.883	0.772
	ANN	0.860	0.134	0.904	0.864	0.861	0.861	0.725
SPAIR-T2W	SVM	0.930	0.067	0.954	0.931	0.929	0.930	0.861
	ANN	0.884	0.114	0.909	0.885	0.883	0.884	0.768

ANN, artificial neural network; SVM, support vector machine; TP, true-positive; FP, false-positive; MCC, Matthews correlation coefficient; CV, cross-validation.

Pairwise comparisons in McNemar’s test showed that no statistical differences were found between SVM and ANN models, implying that the choice of the models was not of substantial importance (P=0.999).

Validation result

Table 10 contains the detailed results of the model validation (testing group). Features extracted from SPAIR T2W sequence (SVM: 0.837, ANN: 0.843) achieved higher predictive accuracy than those derived from T2W sequence (SVM: 0.716, ANN: 0.675), which was consistent with internal validation (training group).

Table 10. Classification results obtained from testing group.

Feature set	Algorithm	TP rate	FP rate	Specificity	Precision	Accuracy	F-measure	MCC
T2W	SVM	0.720	0.298	0.733	0.719	0.716	0.718	0.428
	ANN	0.680	0.329	0.714	0.680	0.675	0.680	0.315
SPAIR-T2W	SVM	0.840	0.165	0.857	0.840	0.837	0.840	0.675
	ANN	0.800	0.157	1.000	0.863	0.843	0.797	0.661

ANN, artificial neural network; SVM, support vector machine; TP, true-positive; FP, false-positive; MCC, Matthews correlation coefficient.

Discussion

Prediction of ESCC response to CRT before treatment initiation plays an important role in patient clinical management and allow for early modification of the treatment protocol in nonresponders. Within this context, prior works have documented the value of medical images in prediction of esophageal tumor response. For example, a meta-analysis suggested that a 50% reduction in SUV_mean or SUV_max between pretreatment PET and a PET obtained within the first two weeks of nCRT was the optimal condition for the prediction of response to esophageal cancer (30). Recent studies also showed the DWI provides complementary information for tumor regression in response to CRT besides measuring the changes in tumor size on anatomical MRI and more pilot studies are currently ongoing (8,31,32). De Cobelli et al. (32) showed that the differences between pre- and post-treatment ADC_median (or ΔADC) values were negatively correlated with pathological response (tumor regression grade, TRG).

More recently, there have been efforts reporting on the value of radiomic analysis for the prediction of treatment response in patient with esophageal cancer. Tixier et al. (24) showed that tumor radiomic features on baseline ¹⁸F-FDG PET was associated with clinical response to definitive CRT. In their study, several features (i.e., homogeneity, entropy, intensity variability, and size-zone variability) achieved superior discriminatory ability (AUCs, 0.82–0.89) over any SUV measurement (AUCs, 0.59–0.70). Tan et al. (33) found that changes of ¹⁸F-FDG PET-based radiomic features over treatment (Δfeatures) appeared more predictive of tumor response than pre- or post-treatment alone. While PET is expensive and time-consuming, CT is still the main imaging method in routine clinical practice. Yip et al. (34) determined the associations of contrast-enhanced CT radiomic features (pre- and post-treatment SD) with pathologic response. However, there has been no report to our knowledge to investigate the clinical potential of radiomic analysis using anatomical MRI (T2W and SPAIR T2W) for predicting the therapeutic response of ESCC. Previous studies have demonstrated that radiomic analysis of T2W MRI could predict the treatment response of patients with rectal cancer (35) and radiomic analysis based on T1W, T2W, and DWI could serve as imaging biomarkers of tumor response to CRT in nasopharyngeal carcinoma (23). To our knowledge, our study is the first to investigate the feasibility of radiomic analysis using T2W- and SPAIR T2W-MRI combined with supervised machine-learning algorithms for the prediction of treatment response to CRT in ESCC patients.

In present work, we found that both two radiomic feature subsets (15 features derived from T2W and 18 features derived from SPAIR T2W sequence) were able to discriminate between responders and nonresponders before CRT initiation with AUC values range from 0.640 to 0.757. The radiomic features were mathematical measurements calculated by depending on the arrangement of pixels, which could characterize the regularity of pixel distribution within the tumor space (36). The distribution of voxels in heterogeneous tumors showed more irregular than that in homogeneous tumors (37). In other words, tumor heterogeneity were correlated with tumor proliferation, cellularity, necrosis, and hypoxia (38) and may be related to poor response and worse prognosis. Tumor response therefore could be predicted, as it is closely associated with tumor heterogeneity. The above results showed that radiomic analysis of anatomical MRI (T2W- and SPAIR T2W) held great potential in differentiating different tumor response in patients with ESCC.

In addition, we analyzed the predictive performance of radiomic feature further by modeling using SVM and ANN algorithm. Zhang et al. (39) showed that radiomic features combined with SVM algorithm achieved higher accuracy in prediction of tumor response to CRT in patients with esophageal cancer. However, their study lacked a validation set (testing group) to evaluate real performance of the classifier. For each sequence, in present work, SVM and ANN were constructed on the training group (n=43) and then validated on the testing group (n=25). Additionally, in order to minimize the risk of modeling over-fitting and bias, we followed a range of robust processing methods: feature reproducibility assessment and dimensionality reduction, wrapper-based feature selection, and models constructed with 10-fold CV. After a series of above-mentioned processing, both two models based on different sequence (T2W and SPAIR T2W) achieved higher predictive accuracy (range, 0.861–0.929). These results might be attributed to the ability of radiomic analysis to indirectly capture the phenotypic information of tumors, including parameters not easily visible and quantifiable by simple visual analysis, which may be linked to tumor response. As for classifiers, SVM (T2W: 0.893; SPAIR T2W: 0.929) achieved better predictive performance than that of ANN (T2W: 0.861; SPAIR T2W: 0.883), although the difference was not statistically significant. Overall, our study indicates that radiomic models appear to be effective in prediction of tumor response and may facilitate better clinical decision making, particularly in the case of patients with ESCC.

However, some limitations are worth noting. Compared to 3D radiomic analysis methods, the only use of 2D slice has some limitations, as it is not sufficient to capture inter-slice features. Because of the larger slice thickness and gap, our study is inappropriate for 3D radiomic analysis method. In addition, due to relatively small sample size of patients and the retrospective nature of this study, our results should be validated in multiple centers with a larger and prospective patient cohort in the future.

In conclusion, combined with supervised machine learning algorithm, radiomic features extracted from T2W- and SPAIR T2W-MRI can be used as an adjunct tool for clinical practice in prediction of treatment response to CRT in patients with ESCC.

Table S1. Feature type and associated features.

Feature type	Methods	Feature name
Histogram-based	IH	Interquartile range (IQR)
		Mean absolute deviation (MAD_1)
		Median absolute deviation (MAD_2)
		Kurtosis
		Skewness
	GH	Interquartile range (IQR)
		Mean absolute deviation (MAD_1)
		Median absolute deviation (MAD_2)
		Kurtosis
		Skewness
Texture-based	GLCM (0°, 45°, 90°, 135°, and mean)	Energy
		Entropy
		Contrast
		Correlation
		Homogeneity
		Variance
		Cluster tendency (CT)
		Cluster shade (CS)
		Inverse difference moment (IDM)
		Inverse variance (IV)
	GLRLM (0°, 90°, and mean)	Short run emphasis (SRE)
		Long run emphasis (LRE)
		Gray-level non-uniformity (GLN)
		Run-level non-uniformity (GLN)
		Low gray-level run emphasis (LGRE)
		High Gray-Level Run Emphasis (HGRE)
		Short run low gray-level emphasis (SRLGE)
		Short run high gray-level emphasis (SRHGE)
		Long run low gray-level emphasis (LRLGE)
		Long run high gray-level emphasis (LRHGE)
		Run percentage (RP)
	NIDM	Busyness
		Coarseness
		Complexity
		Contrast
		Texture strength (TS)
Transform-based	GWTF (5 scales, 8 orientations)	40 mean square energy (MSE)
		40 mean square amplitude (MSA)

IH, intensity histogram; GH, gradient-orientation histogram; GLCM, gray-level co-occurrence matrix; GLRLM, gray-level run-length matrix; NIDM, neighborhood intensity difference matrix; GWTF, Gabor wavelet transform.

Ethical Statement: This study was approved by the Nanjing Drum Tower Hospital’s ethics committee (No. 2015-063-09) and informed consent from the patients was provided.