Modeling Pathologic Response of Esophageal Cancer to Chemoradiotherapy Using Spatial-Temporal ¹⁸F-FDG PET Features, Clinical Parameters, and Demographics

Abstract

Purpose

To construct predictive models using comprehensive tumor features for the evaluation of tumor response to neoadjuvant chemoradiotherapy (CRT) in patients with esophageal cancer.

Methods and Materials

This study included 20 patients who underwent trimodality therapy (CRT + surgery) and had ¹⁸F-FDG PET/CT scans both before and after CRT. Four groups of tumor features were examined: (1) conventional PET/CT response measures (SUV_max, tumor diameter, etc.); (2) clinical parameters (TNM stage, histology, etc.) and demographics; (3) spatial-temporal PET features, which characterize tumor SUV intensity distribution, spatial patterns, geometry, and associated changes resulting from CRT; and (4) all features combined. An optimal feature set was identified with recursive feature selection and cross-validations. Support vector machine (SVM) and logistic regression (LR) models were constructed for prediction of pathologic tumor response to CRT, using cross-validations to avoid model over-fitting. Prediction accuracy was assessed via area under the receiver operating characteristic curve (AUC), and precision was evaluated via confidence intervals (CIs) of AUC.

Results

When applied to the 4 groups of tumor features, the LR model achieved AUCs (95% CI) of 0.57 (0.10), 0.73 (0.07), 0.90 (0.06), and 0.90 (0.06). The SVM model achieved AUCs (95% CI) of 0.56 (0.07), 0.60 (0.06), 0.94 (0.02), and 1.00 (no misclassifications). Using spatial–temporal PET features combined with conventional PET/CT measures and clinical parameters, the SVM model achieved very high accuracy (AUC 1.00) and precision (no misclassifications), significantly better than using conventional PET/CT measures or clinical parameters and demographics alone. For groups with a large number of tumor features (groups 3 and 4), the SVM model achieved significantly higher accuracy than the LR model,

Conclusions

The SVM model using all features including spatial–temporal PET features accurately and precisely predicted pathologic tumor response to CRT in esophageal cancer.

INTRODUCTION

Esophageal cancer remains one of the most lethal malignancies, with a 5-year relative survival rate of only 17% (1) despite continued advances in therapy. In the United States, it is estimated that 17,460 patients were diagnosed with esophageal cancer and 15,070 died from the disease in 2012 (1). The preferred primary treatment strategy for locally advanced esophageal cancer has been transitioning from surgery (esophagectomy) to trimodality therapy, which consists of concurrent neoadjuvant chemoradiotherapy (CRT) followed by surgery (2). Recently, it was suggested that not all patients benefit from surgery after induction CRT and that definitive CRT (CRT alone) could also become an option (3). Evidence suggests that surgery after CRT can significantly improve local control (4, 5). These improvements in local control, however, have been tempered by the increased mortality (9%–12%) and morbidity (30%) compared to CRT alone (mortality, 0.8%–3.5%). Several studies have shown that tumor response to CRT remains an important predictor of both local control and overall survival (3–5). Complete responders to CRT appear to have superior outcomes, regardless of whether they undergo surgical resection. These data also support that the addition of resection can improve outcomes for patients who are discovered to have residual tumor following completion of CRT. Given the added mortality and morbidity of surgery after CRT, as well as the high local failure rate for CRT alone, it is critical to accurately identify patients who respond to CRT so that surgery may be safely deferred. It is equally important to accurately identify patients who do not respond to CRT so that early surgical salvage can be initiated.

Recent studies have emerged suggesting that spatial PET/CT features, including tumor volume (6), tumor shape (7), total glycolytic volume (8), and spatial patterns (texture features) (9), are more informative than the traditional response measure with maximum standardized uptake values (SUV_max) in various tumors. The authors demonstrated that comprehensive spatial–temporal ¹⁸F-FDG PET features were useful predictors of pathologic tumor response to CRT in esophageal cancer (10). The diversity of the new features suggest that it would be advantageous to combine multiple features in evaluation of tumor response (11) instead of traditional PET response criteria that are based on cutoff values of a single measure (8, 12). The objective of this study is to construct sophisticated tumor response models using comprehensive tumor features to accurately and precisely predict pathologic tumor response to CRT in patients with esophageal cancer.

MATERIALS AND METHODS

Patients

This retrospective study was approved by the IRB. The cohort included 20 consecutive patients (median age, 64 years) with esophageal cancer, who underwent trimodality therapy from 2006 to 2009 and had PET/CT scans both before and after CRT (Table 1). Staging was according to AJCC Cancer Staging Manual sixth edition (13), where M1a is extensive local–regional lymph node disease without distant metastasis.

TABLE 1.

Patient Characteristics (n = 20)

Characteristic	No. of patients
Sex
Male	18
Female	2
Primary site
Proximal	0
Distal	20
Mid	0
Throughout	0
Histology
Squamous cell carcinoma	3
Adenocarcinoma	17
Histologic grade
Well differentiated	3
Moderately differentiated	10
Poorly differentiated	5
Unknown	2
TNM stage
T1	0
T2	2
T3	18
T4	0
N0	6
N1	14
M0	18
M1a	2
Pathologic response
Pathologic complete response	5
Microscopic residual disease	4
Gross residual disease	11

PET/CT Imaging, Chemoradiotherapy

Pre-CRT PET/CT imaging was performed 3–5 weeks before the beginning of CRT. Post-CRT PET/CT was performed 4–6 weeks after completion of CRT but before surgery. All PET/CT studies were acquired on the same scanner following an institutional standard protocol (10). All patients were treated with external-beam radiotherapy to the same total dose of 50.4 Gy with concurrent Cisplatin or Carboplatin chemotherapy (14).

Pathologic Assessment

Surgical resection was performed 1–7 weeks after the post-CRT PET/CT, and 6–13 weeks after CRT. The resected surgical specimen was submitted to the same pathologist for evaluation. The specimen was microscopically examined, and semi-quantitatively categorized into 1 of 3 groups: pathologic complete response (pCR), microscopic residual disease (mRD), or gross residual disease (gRD), according to the amount of residual viable carcinoma observed in relation to areas of fibrosis (15). In this study, the primary tumor alone was considered, and both pCR and mRD were considered as “responders,” because they have been shown to have similar survival rates (14), while gRD was considered as “non-responder”.

Spatial–Temporal PET Features Extraction

Details of the spatial–temporal PET feature extraction have been reported (10). Post-CRT PET/CT scans were rigidly registered to pre-CRT PET/CT scans. A tumor volume of interest (VOI) was semi-automatically delineated on the PET image using a region-growing method with a threshold of SUV ≥ 2.5. The result was reviewed and manually edited by a radiologist. The resulting VOI represented the entire hypermetabolic tumor volume and was denoted as VOI_SUV_2.5. Another VOI_SUV_peak was defined as the 3×3×3-voxel cube centered at the SUV_max point, representing the peak metabolically active part of the tumor. Finally, comprehensive spatial–temporal PET features were extracted to characterize the tumor SUV intensity distribution, spatial variations (texture), geometry, and their associated changes resulting from CRT. For each image and each VOI, 9 intensity features (25), 8 Haralick texture features (26), 15 geometry features (27, 28), and 1 volume-intensity feature were extracted (Table 2). After incorporating temporal changes and excluding quantitatively identical features, a total of 137 features were obtained for each tumor.

TABLE 2.

Spatial-Temporal ¹⁸F FDG-PET Features

Intensity features

Minimum

Maximum

Mean

Standard deviation

Sum

Median

Skewness

Kurtosis

Variance

Texture features

Energy

Entropy

Correlation

Inverse difference moment

Inertia

Cluster shade

Cluster prominence

Haralick correlation

Geometry features

Volume

Major diameter

Minor diameter

Eccentricity

Elongation

Oriented bounding box volume

Bounding box volume

Roundness

Region ratio

Orientation

Feret diameter

Number of lines

Perimeter

Physical size

Flatness

Geometry–intensity feature

Total glycolytic volume

Clinical Parameters and Demographic Features

Sixteen clinical parameters and demographic features were extracted from patients’ charts. Clinical parameters included differentiation, stage, T stage, N stage, M stage, distant metastasis, type of chemotherapy, radiotherapy dose, treatment with concomitant boost, location of tumor, tumor involves gastroesophageal junction, histology, total extent of disease, and extent of disease >4 cm. Demographic features included age and sex.

Feature Selection and Response Modeling

Four groups of tumor features were examined: (1) 16 conventional PET/CT response measures (Table 3); (2) 16 clinical parameters and demographics as described above; (3) 137 spatial–temporal PET features (Table 2); and (4) all 169 combined features. Figure 1 illustrates the flow of feature selection and model construction, with cross-validations to remove bias in feature selection and to prevent over-fitting of the model, respectively. The pathologic tumor response of the surgical specimen was used as the ground truth. The 20 patients were randomly partitioned into a training set (for constructing the model) and a testing set (for cross-validating the model and assessing its accuracy). 10-, 5- and 2-fold cross-validations (randomly leaving 10%, 20%, and 50% patients out in the testing set) were repeatedly used for both feature selection and model accuracy evaluation. Our goal was to model pathologic tumor response to CRT as a function (f) of each of the 4 groups of tumor features so that:

$$\begin{array}{l}\text{Pathologic}\text{tumor}\text{response}=f(\text{conventional}\text{PET}/\text{CT}\text{measures},\text{or}\text{clinical}\text{parameters}\text{and}\\ \text{demographics},\text{or}\text{spatial}-\text{temporal}\text{PET}\text{features},\text{or}\text{all}\text{combined}\text{features})\end{array}$$ Eq. 1

TABLE 3.

Conventional PET/CT Measures Used in Tumor Response Evaluation^*

SUV features

SUVmean (post, VOI_SUVpeak)

SUVmax (post, VOI_SUV2.5)

SUVmean Decrease (VOI_SUVpeak)

SUVmean (pre, VOI_SUVpeak)

SUVmax (pre, VOI_SUV2.5)

SUVmean ratio (pre, post, VOI_SUVpeak)

SUVmax decrease (pre, post, VOI_SUV2.5)

SUVmax ratio (pre, post, VOI_SUV2.5)

Volume and length features

Major diameter (post, VOI_SUV2.5)

Major diameter Ratio (pre, post, VOI_SUV2.5)

Volume (post, VOI_SUV2.5)

Volume ratio (pre, post, VOI_SUV2.5)

Major diameter (pre, VOI_SUV2.5)

Major diameter decrease (pre, post, VOI_SUV2.5)

Volume (pre, VOI_SUV2.5)

Volume decrease (pre, post, VOI_SUV2.5)

^*Parentheses denote VOI and images from which a feature was extracted.

pre = Pre-CRT SUV, post = Post-CRT SUV, diff = pre-CRT SUV – post-CRT SUV.

FIGURE 1.

Workflow diagram illustrating feature selection and model construction with cross-validations for prediction of tumor response.

Feature Ranking and Selection

Since there are many features, to avoid model over-fitting to the training set, a feature selection process with cross-validation (Fig. 1) was applied first within each feature group for each model. A support vector machine–based feature ranking (SVMFR) method (16) was used to rank all features by recursively removing features and testing the predictive ability of the remaining features for each patient partition. A frequency distribution of the top-ranked 20 features for each of the k partitions was obtained. The optimal feature set was then identified as the 10% most frequently selected features over all k partitions from the frequency distribution. In this manner, the feature selection bias was removed.

Predictive Model Construction

We used 2 machine learning (ML) models, logistic regression (LR) and SVM, to obtain function f in Eq. 1. Figure 1 illustrates the modeling process using SVM with all tumor features as an example.

SVM with sequential minimal optimization

In our SVM, a polynomial kernel was used to transform the input tumor features from x₁, … x_m space into x̃₁, … x̃_m space, so that the 2 classes (“responders” and “non-responders”) became linearly separable (Fig. 1). The hyperplane (or the straight line in the simplified 2D illustration) that represents the largest separation between the closest members (support vectors) of the two classes was determined, providing a classification rule or model y = f (x̃₁,…, x̃_m), which classified a new or testing patient as a “responder” or a “non-responder” based on its y value.

Logistic regression

For comparison, we also implemented the widely used LR model. LR first transformed the response variable into Pr[YES | x₁,…, x_m], a probability variable corresponding to the “responders” given tumor features x₁,…, x_m. A logit transformation, log{Prp[YES | x₁,…, x_m]/(1−Pr[YES | x₁,…, x_m])}, was then applied so that the resulting variable lies between negative infinity and positive infinity. The transformed variable was approximated using a linear function of input tumor features (linear regression). The resulting model was Pr[YES | x₁,…, x_m] = 1/[1+exp(−w₀ −w₁x₁ –…– w_m x_m)] with weights w. The weights were obtained by fitting the model to the training set using maximum log-likelihood estimation.

The outputs of the ML models were the predicted pathologic response represented as a binary variable (yes or no), which corresponded to “responder” or “non-responder”, respectively (Fig. 1). To test over-fitting of each model, 10-, 5- and 2-fold cross-validations were again used in model evaluation (Fig. 1). The model accuracy was calculated as the mean accuracy over all partitions.

Statistical Analysis

The accuracy of using each feature group and each model to predict the pathologic tumor response was quantified using the AUC, defined as the area under receiver operating characteristic (ROC) curves. In addition, the sensitivity and specificity of each model were calculated and compared using the unpaired t test at a significance level of 0.05. Model precision was evaluated with the 95% confidence intervals (CI).

RESULTS

Because LR and SVM are 2 distinct models, our feature selection process resulted in different optimal feature sets for each model (Table 4). The optimal feature set for SVM contained the optimal feature set for LR, except when applied to clinical parameters and demographics, where histology was the only feature selected for SVM. This was in agreement with a larger study of 164 patients by Koshy et al. (14) that showed histology was the most and only predictive individual clinical parameter.

TABLE 4.

Optimal Feature Set Resulting from Feature Selection^*

Metrics/Features	Logistic regression	Support vector machine
Conventional PET/CT measures	Major diameter (pre, VOI_SUV2.5)	Major diameter (pre, VOI_SUV2.5) Volume (pre, VOI_SUV2.5) SUVmax Decline (pre, post, VOI_SUV2.5) SUVmax Ratio (pre, post, VOI_SUV2.5)
Clinical parameters and demographics	Histology Tumor involves gastroesophageal junction Sex T stage	Histology
Spatial–temporal PET features	Inertia (diff, VOI_SUVpeak) Orientation (post, VOI_SUV2.5) Variance (post, VOI_SUVpeak) Cluster prominence (post, VOI_SUVpeak) Inverse difference moment (post, VOI_SUV2.5)	Inertia (diff, VOI_SUVpeak) Orientation (post, VOI_SUV2.5) Variance (post, VOI_SUVpeak) Cluster prominence (post, VOI_SUVpeak) Inverse difference moment (post, VOI_SUV2.5) Energy (post, VOI_SUV2.5) Entropy (post, VOI_SUV2.5) Haralick correlation (pre, VOI_SUV2.5) Median (post, VOI_SUV2.5) Inverse difference moment (Diff, VOI_SUVpeak) Median (post, VOI_SUVpeak) Entropy (diff, VOI_SUVpeak) Roundness (post, VOI_SUV2.5) Mean (post, VOI_SUV2.5) Elongation (pre, VOI_SUVpeak) Total glycolytic volume (diff, VOI_SUV2.5)
All features	Orientation (post, VOI_SUV2.5) Tumor involves gastroesophageal junction Inertia (diff, VOI_SUVpeak) Energy (post, VOI_SUV2.5) Entropy (post, VOI_SUV2.5) Skewness (diff, VOI_SUVpeak)	Orientation (post, VOI_SUV2.5) Tumor involves gastroesophageal junction Inertia (diff, VOI_SUVpeak) Energy (post, VOI_SUV2.5) Entropy (post, VOI_SUV2.5) Skewness (diff, VOI_SUVpeak) Inverse difference moment (post, VOI_SUV2.5) Cluster prominence (post, VOI_SUVpeak) Inverse difference moment (diff, VOI_SUVpeak) Inertia (post, VOI_SUV2.5) Volume (post, VOI_SUV2.5) Elongation (pre, VOI_SUVpeak) T stage Variance (post, VOI_SUVpeak) Haralick correlation (pre, VOI_SUV2.5) Median (post, VOI_SUV2.5) Total glycolytic volume (diff, VOI_SUV2.5)

^*Parentheses denote VOI and images from which a feature was extracted.

pre = Pre-CRT SUV, post = Post-CRT SUV, diff = pre-CRT SUV – post-CRT SUV.

Figure 2 shows the model accuracy (AUC) and precision (95% CI) obtained from repeating the 10-, 5-, and 2-fold cross-validations. The best prediction was obtained using the SVM model with 17 features from all combined features (SVM_all). All patients within the testing set were correctly classified during the repetition of 10-fold cross-validations, resulting in a mean AUC of 1.00 (100% sensitivity, 100% specificity). Table 4 indicates that SVM_all contained 1 conventional PET/CT measure, “residual metabolic tumor volume (i.e., SUV ≥ 2.5) post-CRT”; 2 clinical parameters, “whether tumor involves gastroesophageal junction” and “T stage”; and 14 spatial–temporal PET (3 intensity, 8 texture, 2 geometry, and 1 volume-intensity) features, suggesting that all 3 groups of tumor features and all 4 categories of spatial-temporal PET features contained useful predictors of response. The model performance was stable when leaving more patients out with 5- and 2-fold cross-validations compared to 10-fold cross-validation. Only a small reduction in mean AUC (from 1.00 to 0.99 and 0.92) was observed.

FIGURE 2.

Response modeling accuracy and reliability: mean AUC, sensitivity, and specificity of 10-, 5-, 2-fold cross-validations. Error bars indicate 95% confidence intervals.

Figure 2 also shows the sensitivity and specificity obtained from each model using different groups of features. When the SVM model was used with 10-fold cross validation, significantly higher sensitivity was achieved by using all features including spatial–temporal PET features (100%) than by using conventional PET/CT measures (60%) or clinical parameters and demographics (70%) alone (P < 0.001). Significantly higher specificity (P < 0.001) was achieved as well. Similar results were obtained when the LR model was used (92% sensitivity, 94% specificity; both P < 0.001).

For 10-fold cross-validation, the differences between SVM and LR models were not significant when using any of the four groups of features (P > 0.06). However, when using selected spatial–temporal PET features or using all features in 5- and 2-fold cross-validations, the SVM models demonstrated significantly better results than the LR models (P < 0.0001 and P < 0.0002, respectively).

DISCUSSION

¹⁸F-FDG PET has shown promising results in predicting pathologic response to CRT and long-term prognosis in esophageal cancer (12, 17). Westerterp et al. (18) and Swisher et al. (19) showed that PET had the highest accuracy (76%) among PET, endoscopic ultrasonography, and CT for predicting pathologic response to CRT with sensitivity ranges of 71%–100% and specificity ranges of 55%–100%. Levien et al. (20) showed that PET can be useful for predicting pathologic response with sensitivity of 61.3% and specificity of 60.0%. Reviewing 20 studies, Kwee (12) found that the sensitivity and specificity of PET for predicting pathologic response to neoadjuvant therapy ranged from 33% to 100% and from 30% to 100%, respectively. Although promising, the accuracy of PET is still low and none of these studies achieved both high sensitivity and high specificity.

Almost all published ¹⁸F-FDG PET studies quantify therapeutic response in tumors with SUV_max (21, 22). Changes in SUV_max and sometimes SUV_max before (pre-) or after (post-) CRT are correlated to post-CRT pathologic response or survival. However, SUV_max is a single-point estimate, whereas most solid tumors show significant heterogeneity in both degree and distribution of FDG uptake. Heterogeneity in FDG uptake has been associated with important biological and physiologic parameters (9, 23) and has been shown to be a prognostic factor in many cancers (9, 24). In our previous work, comprehensive spatial–temporal ¹⁸F-FDG PET features were extracted and found to be useful predictors of pathologic tumor response (10).

In this work, a predictive model generally included multiple tumor features. The best model achieved significantly higher accuracy (AUC 1.00) than the best individual response measure (inertia post-CRT, AUC 0.85, (10)). Both prediction accuracy and precision were significantly improved by using selected spatial–temporal PET features instead of conventional PET/CT measures or clinical parameters and demographics. This was the case for both the SVM and LR models. These results suggested that spatial–temporal PET features provided richer information than the other feature groups.

When using the same feature group and comparing the performance of LR and SVM models, the results varied from group to group. SVM achieved significantly higher accuracy than LR when using spatial–temporal PET feature group. The reason is that this group contained more candidate features, whose complimentary relationship for response prediction is hard to identify with LR. On the other hand, SVM has been proven to be able to extract complex relationships among a large number of features (25). Because the candidate feature group of the conventional PET/CT measures or clinical parameters and demographics contained only 16 features and because only 1 or 4 features were selected into the optimal subset, LR resulted in better results than SVM. The reason is that with this small number of selected features it would be difficult for SVM to achieve high accuracy.

When utilizing many candidate features for classification, it is important to identify an optimal, smaller feature set to prevent model over-fitting. Our feature selection process removed redundant features that introduce colinearity and noise into the models (26). Furthermore, feature selection bias was eliminated by collecting frequency distribution of the selected features over all patient partitions. The feature selection resulted in 16 features for the SVM using spatial–temporal PET features alone. When using all combined features, the number of selected features was 17. 1 conventional PET/CT measure and 2 clinical parameters replaced 2 spatial-temporal PET features, resulting in increased prediction accuracy (AUC increased from 0.94 to 1.00). This means features from conventional PET/CT measures and clinical parameters provided some complementary information to the spatial–temporal PET features.

Another important aspect of constructing predictive models is to avoid model over-fitting. To test this, we used different number of patients to train and test our models, namely 10-, 5- and 2-fold cross-validations. When leaving more patients out of training set, the prediction accuracy decreased. However, the AUC was still above 0.90 for SVM_all model suggesting that it was not notably affected by over-fitting. LR model was not as stable as SVM model in this case (AUC dropped from 0.90 to below 0.70).

The authors provided an explanation as to why spatial–temporal PET features were useful predictors of tumor response (10). In addition to the spatial–temporal PET features, 1 conventional PET/CT measure (residual metabolic tumor volume) and 2 clinic parameters (whether tumor involves gastroesophageal junction and T stage) were included in the optimal feature set for the SVM_all model. Evidence suggests that smaller residual metabolic tumor volume (27) and lower T stage (13) are associated with better tumor response. The effects of tumor involvement of the gastroesophageal junction, however, are not yet clear and are under investigation at our institution using a larger dataset.

There were variations in the interval between CRT and surgery (6–13 weeks) and the interval between post-CRT PET/CT and surgery (1–7 weeks). Several studies suggest that in rectal cancer delaying surgery after CRT may increase the pCR rate. However, based on our knowledge, three studies in esophageal cancer suggest that the timing of surgery did not affect the pCR rate (28–30). Examination of our data using Z-test suggested that neither interval affected the pCR rate (both P = 0.09). The results were not affected by variations in the intervals.

One limitation of this proof-of-principle study is that pathologic response in the primary tumor alone was examined. A pCR on the primary tumor site may not completely exclude lymph node invasion. An extension of the study to lymph node is under investigation. Another limitation is that it is a retrospective analysis of a small patient cohort. Although 10-, 5- and 2-fold cross-validations showed that the model was not notably affected by over-fitting, the predictive accuracy and stability of the models should be validated in a larger, independent patient cohort as shown in Fig. 1. Validation is also needed to confirm that the selected features are indeed meaningful measures and important for response evaluation in esophageal cancer. When the model is validated, it can be used to more appropriately select patients for surgery, thus avoiding the mortality and morbidity of surgery in responders for whom surgery can be safely deferred. The methodology can also be applied to evaluate response during CRT, which will provide the opportunity for early adjustments on treatment strategies including: giving a higher dose in definitive CRT to responders, changing the type of chemotherapy or performing surgery earlier to non-responders.

CONCLUSION

The SVM model using all features including spatial–temporal PET features accurately and precisely predicted pathologic tumor response to CRT in 20 patients with esophageal cancer. It has the potential to be used to safely defer surgery or to give a higher dose in definitive CRT for patients who respond to CRT. This will ultimately improve patient’s quality of life while reducing costs.

Summary.

This work constructed sophisticated predictive models using multiple, comprehensive tumor response measures, including conventional PET/CT measures, clinical parameters and demographics, and spatial–temporal PET features. The models achieved very high accuracy (100% sensitivity, 100% specificity) for prediction of pathologic tumor response to chemoradiotherapy in 20 patients with esophageal cancer.