Textural features based enhanced contrast CT images predicts prognosis to concurrent chemoradiotherapy in stage III esophageal squamous cell cancer

2.1 Treatment

Chemotherapy consisting of a Taxol (135 mg/m 2 , D1) plus cisplatin (25 mg/m 2 , D1–3) and Docetaxel (60 mg/m 2 , D1) plus cisplatin (25 mg/m 2 , D1–3) was given concurrently with radiation therapy, for which the median dose was 64 Gy (range: 60–66 Gy), delivered in daily 1.8–2.0 Gy fractions on Monday through Friday. Radiation was delivered by three-dimensional conformal radiation therapy, intensity-modulated radiation therapy. A total of 17 (29.8%) patients also received adjuvant chemotherapy, which included a Taxol plus cisplatin (TP) and Docetaxel plus cisplatin (DP).

2.2 CT image acquisition

All patients underwent standard radiotherapy simulator contrast, material-enhanced CT scans of the thorax and abdomen before treatment. All simulator CT studies in our center were performed with a 16–detector row CT scanner (PHILIPS Healthcare) and the same clinical protocol: 120 kV, 180–280 mA with tube current modulation, a matrix of 512, a field of view of 500 mm, and 3 mm reconstructed section thicknesses after intravenous injection of 60–120 ml of 370 mg/ml (1.5 ml/Kg) iodinated contrast agent at 1.8–2 mL/sec. The thoracic images (from the sternal notch to below the hemidiaphragm) were obtained in the arterial phase 0–5 seconds after contrast agent administration. Abdominal images (from above the hemidiaphragm to the iliac crest) were obtained in the portal venous phase 60 seconds after contrast administration. Contrast CT image sets obtained from all 57 patients before treatment and RT 40 Gy were available in institutional archives for analysis.

2.3 Texture analysis

Whole-tumor texture was analyzed using IBEX (Imaging Biomarker Explorer) software version 1.0 β [12]. Two radiation oncologists, Li and Wang with more than 10 years of CT experience and 1 year of texture analysis experience, delineated the whole-tumor volume by using region of interest (ROI) analysis on contiguous 3-mm CT images. Esophageal stents, nasogastric tubes, intraluminal air, and oral contrast material were excluded from the ROIs (Fig. 1). This process was refined by an additional thresholding procedure that excluded any pixels with attenuation of less than - 50 HU.

OPEN IN VIEWER

Texture analysis was performed in a two-stage process that involved image filtration (image transformation) followed by statistical quantification of texture (Fig. 1). Image filtration involved using a Laplacian of Gaussian spatial band-pass filter to produce a series of derived images highlighting features at different anatomic spatial scales ranging from fine to coarse texture for the ROI. The different filter values used in this study were 1.0 (highlighting fine textures, which may enhance tissue parenchymal features), 1.5–2.0 (highlighting medium textures), and 2.5 (highlighting coarse textures, which may enhance vascular features). The filter widths represented by fine, medium, and coarse textures were 4, 6–8, and 10 pixels, respectively, with four categories of quantitative image features (Table 2): intensity direct and histogram (global entropy, global maximum, global mean gray-level intensity, global standard deviation of the histogram [Global Std], global uniformity, kurtosis, skewness).

The Kaplan-Meier method was used to construct OS and PFS curves, and the log-rank test was used to determine the statistical significance of differences. PFS was computed from the time of treatment to the time of tumor progression, or death without evidence of disease progression, at which point the data were censored. OS was calculated from the time of treatment to the time of death from any cause, or to the time of last follow-up, at which point the data were censored. The proportional changes in each texture parameter after CRT was calculated as the differences between the texture parameter pre-treatment and that post-treatment (radiation dose 40 Gy) with regard to the baseline texture value. Changes in textures ( C t ) after CRT were defined as the following equation:

C t = X 1 - X 2 X 1 × 100 %

where X 1 is the pre-texture parameters, X 2 is the post-texture parameters. Texture parameters at each filter level and their proportional changes after CRT were correlated with survival time. Univariate analyses using the Cox proportional hazards model were performed on each texture parameter as a continuous variable for all four filter values. Median value for each texture parameter was used as threshold levels to divide patients for Kaplan-Meier survival analysis, with comparisons between groups by the log-rank test.

The texture parameters with 0.05 in both Cox model and Kaplan-Meier analysis were characterized as significant univariate prognostic factors. The significant texture parameters and morphologic variables in univariate analyses were used in multivariate Cox proportional hazards model. The optimal Cox model (best combination of variables) was selected by a step-wise procedure with exclusion and inclusion of candidate variables, based on the Akaike information criterion (AIC). A risk score of prognosis was predicted by the optimal Cox model. Time-independent receiver operating characteristic (ROC) curve analysis was performed to compare the accuracy of the optimal Cox model to those of the texture parameters alone, or the morphologic assessment alone, in assessing the chance of survival. Areas under the ROC curve (AUCs) of each group were calculated and compared by using the DeLong test. For Kaplan-Meier curves, patients were divided into two subgroups by the median value of risk score or the texture parameters. All the statistical methods were performed by two sides, with 0.05 as significance cut-off. These statistical analyses were performed by the R software.

3.1 Patients

All patients had stage III ESCC disease. Most patients (48/57, 84.2%) had primary tumors in the upper and middle thirds of the esophagus (Table 1). Most patients (36/57, 63.2%) received concurrent DP chemotherapy at weeks 1 and 5. Other patients received concurrent TP chemotherapy. The median times from the pre-treatment CT study to the start of radiation therapy and from the completion of 40 Gy radiation therapy to the post-treatment CT study were 6 days (range, 3–10 days) and 30 days (range, 28–35 days), respectively. Median follow-up for the entire cohort was 32.0 months (range, 7–60 months). The median 1 and 3 years overall survival (OS) was 20.2 months, 75.4% and 32.1%. For clinical factors, in the univariate analysis performed using the log-rank test (Table 1), chemotherapy regimen ( P = 0.024 HR = 0.463, 95% confidence interval [CI] 0.237–0.905), pre-volume ( P = 0.039 HR = 2.033, 95% CI 1.037–3.988), and post-volume ( P = 0.029 HR = 2.116, 95% CI 1.079–4.152) were found to be significant prognostic factors.

3.2 Primary tumor CT features for survival prediction

For primary tumor, in the univariate analysis performed using the Cox proportional hazards model (Table 2), pre-treatment fine, medium, and coarse texture global entropies, global mean, global Std, and global uniformity as continuous variables were found to be significant prognostic factors. For filter value 1.0, the global Std P = 0.0024, the hazard ratio [HR] was 0.602 (95% [CI]: 0.434–0.836); value 1.5, the HR was 0.311 (95% [CI] 0.141–0.687; P = 0.0039). At value 2.0, the HR was 0.185 (95% [CI] 0.051–0.675; P = 0.0106).

OPEN IN VIEWER

In the univariate analysis performed using the log-rank test (Table 2), the cutoff value median of pretreatment fine, medium 1–2, and coarse texture global entropies, global mean, global Std, and global uniformity were found to be significant prognostic factors for filter value 1.0, global entropies (cutoff = 0.4942, P = 0.0119), global mean (cutoff = 2.5622, P = 0.088), global Std (cutoff = 4.3551, P = 0.0209), and global uniformity, (cutoff = 0.8133, P = 0.0284). Other filter value is shown in Table 2.

For metastasis lymph nodes, in the univariate analysis performed by using the Cox proportional hazards model (Table 2), pre-treatment fine texture global entropy (HR = 0.186, 95% CI: 0.0567–0.608, P = 0.0054), global mean (HR = 0.752, 95% CI: 0.612–0.924, P = 0.0068), global Std (HR = 0.662, 95% CI: 0.662–0.943, P = 0.0091), global uniformity (HR = 34.844, 95% CI: 2.655–457.293, P = 0.0069), and skewness (HR = 1.609, 95% CI: 1.079–2.402, P = 0.0197) as continuous variables were found to be significant prognostic factors. Posttreatment fine texture global entropy and mean can be prognostic factors. In the univariate analysis performed using the log-rank test (Table 2), the cutoff value median of posttreatment fine global mean was found to be a significant prognostic factor ( P = 0.0085).

3.4 Changes in CT textures with CRT and survival prediction

In the univariate analysis performed using the Cox proportional hazards model (Table 2), coarse texture (value 2.5), primary tumor skewness and kurtosis changes as continuous variables were found to be significant prognostic factors: (HR = 0.8011, 95% CI: 0.6566–0.9774, P = 0.0289) and (HR = 0.8889, 95% CI: 0.8211–0.9624, P = 0.0037). Metastasis node-only global entropy change (fine texture value 1.0) was a prognostic factor (HR = 0.8146, 95% CI: 0.7210–0.9204, P = 0.0009).

3.5 Survival models

In the whole group, four parameters (post-volume, entropy, mean, global Std) as continuous variables were found to be significantly correlated with OS in univariate Cox analysis, which was consistent with the Kaplan-Meier result (Table 2). A multivariate Cox prediction model was constructed by the four parameters and was optimized in feature selection by a step-wise exclusion and inclusion of variables. The final optimal prediction model was constructed with a combination of two parameters of post-volume (Fig. 3A) and global Std (Fig. 3B, Table 3). Time-independent ROC analysis revealed that the combined model performs a better predictive value for OS than global Std in the time window of 12 months, and then post-volume in the time window of 24 months. Using 12 months as the endpoint, the combined model got an AUC of 0.79, which was higher than that of global Std (AUC = 0.71, p = 0.107) in assessing the chance of patient death (Fig. 2). Using 24 months as the endpoint, the combined model got an AUC of 0.79, which was higher than that of post-volume (AUC = 0.69, P = 0.107) and global Std (AUC = 0.77, P = 0.607) (Fig. 2). The patients were divided into high-risk and low-risk groups by the median risk score of the combined Cox model. Kaplan-Meier curves show that the patients in the high-risk group had significantly worse OS (log-rank test, P = 0.0009) and PFS (log-rank test, P = 0.0019) than those in the low-risk group (Fig. 3C).

OPEN IN VIEWER

For patients who underwent DP, low risk increased OS over that of high risk ( P = 0.003). For patients who underwent TP, low risk increased OS over that of high risk ( P = 0.010). Survival models using texture parameters and morphologic response as assessed by volume did perform better than standard morphologic assessment.