Predicting the Risk of Thymic Tumors Using Texture Analysis of Contrast-Enhanced Chest Computed Tomography

Objective

This study aimed to explore the value of contrast-enhanced computed tomography texture features for predicting the risk of malignant thymic epithelial tumor.

Methods

Data of 97 patients with pathologically confirmed thymic epithelial tumors treated at in our hospital from March 2015 to October 2021 were retrospectively analyzed. Based on the World Health Organization classification of thymic epithelial tumors, patients were divided into a high-risk group (types B2, B3, and C; n = 45) and a low-risk group (types A, AB, and B1; n = 52). Texture analysis was performed using a first-order, gray-level histogram method. Six features were evaluated: mean, variance, skewness, kurtosis, energy, and entropy. The association between contrast-enhanced computed tomography texture features and the risk of malignancy in thymic epithelial tumors was analyzed. The predictive thresholds of predictive texture features were determined by receiver operating characteristics analysis.

Results

The mean, skewness, and entropy were significantly greater in the high-risk group than in the low-risk group (P < 0.05); however, variance, kurtosis, and energy were comparable in the two groups (P > 0.05). The area under curve of mean, skewness, and entropy was 0.670, 0.760, and 0.880, respectively. The optimal cutoff value of entropy for predicting risk of malignancy was 7.74, with sensitivity, specificity, and accuracy of 80.0%, 80.0%, and 75%, respectively

Conclusions

Contrast-enhanced computed tomography texture features, especially entropy, may be a useful tool to predict the risk of malignancy in thymic epithelial tumors.

Thymic epithelial tumors are the most common tumors of the mediastinum and are known to have potential for malignant behavior.¹ The 2004 World Health Organization (WHO) criteria classify thymic epithelial tumors according to risk of malignancy as low risk (types A, AB, and B1) and high risk (types B2, B3, and C).^2,3 Accurate preoperative evaluation of risk can help in selecting the appropriate treatment and predicting prognosis.⁴

Computed tomography (CT)–guided biopsy or thoracoscopic biopsy can clarify the pathological nature of tumors but cannot fully reflect the heterogeneity of the tumors or provide information, on invasion of surrounding structures or lymph node metastasis.⁵ Chest CT is the most commonly used imaging modality for diagnosis mediastinal lesions. Although contrast-enhanced computed tomography (CECT) can help assess risk of malignancy for large thymic epithelial tumors, it is not reliable for small tumors with regular well-defined margins.^6,7

Texture analysis, a branch of radiomics, is a potentially useful method for identifying malignancy risk in tumors. It is based on variations in gray-level patterns in medical images and is not influenced by clinical factors.^8,9 Quantitative CT texture analysis has been applied for grading, staging, and prognosis evaluation of tumors of lung,¹⁰ kidney,^11,12 liver,¹³ prostate,¹⁴ stomach,¹⁵ brain,¹⁶ head and neck,^17,18 breast,¹⁹ bladder,²⁰ colon and rectum,²¹ and soft tissues.²² It has also been used for grading and staging of thymic epithelial tumors and identifying histological subtypes.^23–28

The aim of this study was to evaluate whether texture features on CECT images of thymic epithelial tumors could be used to predict the risk of malignant behavior.

MATERIALS AND METHODS

Study Population

The sample for this retrospective study was selected from among 120 patients (55 female, 65 male; age range, 46–71 years; mean age, 47.3 ± 13.1 years) with thymic tumors who underwent chest CECT before thoracoscopic thymectomy or open thymectomy at our hospital during March 2015 to October 2021. Patients were eligible for inclusion if they (1) had pathologically confirmed thymic epithelial tumors and (2) had undergone CECT before surgery. Patients who had not undergone CECT (n = 12), with severe image artifacts (n = 4), and without pathologic confirmation of diagnosis (n = 7) were excluded. A total of 97 patients (49 female, 48 male; age range, 43–65 years; mean age, 49.5 ± 10.2 years) met the eligibility criteria. All patients were of Mongoloid or Han ethnicity.

This retrospective study was approved by the institutional review board and was conducted in accordance with the Declaration of Helsinki (sixth edition). The board waived the need for written informed consent.

CT Examination

Chest CECT was performed using a 64-channel multidetector CT scanner (GE Discovery CT750 HD; GE HealthCare), with the patient lying supine in the middle of the bed with arms extended over the head. Scanning range was from the thoracic inlet to the diaphragm. The CECT scanning parameters were as follows: tube voltage, 120 kVp; tube current, 200 mA; slice thickness, 5 mm; slice spacing, 5 mm; pitch, 0.75; field of view, 300 × 300 mm; and matrix, 512 × 512. Images were acquired 25 seconds after intravenous injection of 1 mg/kg of iopromide (370 mg/mL; Bayer Medical and Health Care Co, Ltd, Guangzhou, China). A total of 80 to 100 mL was injected per patient, with flow rate of 3.0 mL/s.

Imaging Analysis

Axial CECT images were imported into United Imaging Radiomics Research Platform software v1.0 (Shanghai United Imaging Medical Technology Co, Ltd, China) in Digital Imaging and Communications in Medicine format. Texture analysis was performed using a first-order gray-level histogram method. Six features were evaluated: mean, variance, skewness, kurtosis, energy, and entropy. Two radiologists, each with 10 years of experience in thoracic imaging and blinded to the pathological results, manually delineated a region of interest around the tumor using the United Imaging Radiomics Research Platform software and then independently performed texture analysis. Differences in interpretation were settled by discussion until consensus was reached.

Statistical Analysis

SPSS 20.0 (IBM Corp, Armonk, NY) was used for statistical analysis. Data were expressed as mean ± SD and compared between the high-risk and low-risk groups using the independent-samples test and Mann-Whitney U test. Diagnostic accuracy was determined using the area under the curve, and the cutoff value, sensitivity, and specificity were calculated. Interobserver agreement in evaluation of CECT texture features was assessed using the intraclass correlation coefficient. P < 0.05 was considered statistically significant.

RESULTS

Clinical Characteristics

Among the 97 patients, 32 underwent thoracoscopic thymectomy and 65 underwent open thymectomy. The tumor types included type A (n = 17), type AB (n = 7), type B1 (n = 28), type B2 (n = 26), type B3 (n = 14), and type C (n = 5). Thus, by the 2004 WHO criteria, 45 patients were classified as high risk and 52 as low risk (Fig. 1).

FIGURE 1.

Segmentation of thymomas. A, Contrast-enhanced CT images of a 42-year-old man with WHO type A thymoma; (B) CECT images of a 60-year-old woman with WHO type B2 thymoma. This figure can be viewed online in color at www.jcat.org.

Sex distribution and age were comparable between the high-risk group and the low-risk group. Tumor volume was significantly larger in the high-risk group (Table 1).

TABLE 1.

Characteristics of Patients (n = 97)

Who Type	Number	Race/Ethnicity	Sex (Male/Female)	Age, Mean ± SD, y	Volume, cm3
Low risk	52	Mongoloid/Han	24/28	50.12 ± 8.56	49.22 ± 76.08
Type A	17	Mongoloid/Han	7/10	48.33 ± 7.36	52.47 ± 86.23
Type AB	7	Mongoloid/Han	5/2	50.24 ± 9.16	54.52 ± 92.41
Type B1	28	Mongoloid/Han	12/16	51.18 ± 8.38	53.69 ± 87.72
High risk	45	Mongoloid/Han	24/21	51.39 ± 9.71	98.76 ± 187.52
Type B2	26	Mongoloid/Han	15/11	49.42 ± 9.37	112.21 ± 197.34
Type B3	14	Mongoloid/Han	6/8	51.12 ± 7.33	119.21 ± 205.22
Type C	5	Mongoloid/Han	3/2	51.45 ± 9.77	127.21 ± 201.33
P			0.072	0.199	0.009

Texture Features and Diagnostic Performance

Interobserver agreement in analysis of texture was moderate (intraclass correlation coefficient, 0.886–0.907). Mean, skewness, and entropy were all significantly greater in the high-risk group than in the low-risk group (P = 0.016, P = 0.029, P = 0.005, respectively; Table 2). Variance, kurtosis, and energy were comparable between the two groups (P = 0.457, P = 0.221, P = 0.362, respectively; Table 2). The area under curve for mean, skewness, and entropy was 0.670, 0.760, and 0.880, respectively. Receiver operating characteristics analysis showed that the optimal cutoff value for entropy was 7.74, providing sensitivity, specificity, and accuracy of 80.0%, 80.0%, and 75%, respectively (Table 3, Fig. 2).

TABLE 2.

CECT Texture Features for Low-Risk Versus High-Risk Group

Texture Features	Low Risk (n = 52)	High Risk (n = 45)	P
Mean	82.54 ± 8.22	93.22 ± 7.52	0.016*
Variance	21.38 ± 2.37	21.57 ± 3.52	0.457
Skewness	0.67 ± 0.03	0.73 ± 0.07	0.029*
Kurtosis	0.63 ± 0.06	0.65 ± 0.03	0.221
Energy	0.05 ± 0.02	0.06 ± 0.03	0.362
Entropy	7.42 ± 0.68	8.76 ± 1.44	0.005*

TABLE 3.

Diagnostic Efficiency of CECT Texture Features

Texture Features	AUC (95%CI)	P	Sensitivity	Specificity	Accuracy	Cutoff
Mean	0.670 (0.421–0.919)	0.199	70%	60%	65%	90.91
Skewness	0.760 (0.550–0.970)	0.049	70%	70%	70%	0.715
Entropy	0.880 (0.705–1.000)	0.004	80%	80%	75%	7.74

AUC, area under curve.

FIGURE 2.

Receiver operating characteristics curves for differentiation of low-risk from high-risk thymic epithelial tumors based on the mean, skewness, and entropy. ROC, receiver operating characteristic. This figure can be viewed online in color at www.jcat.org.

DISCUSSION

In this study, a first-order gray-level histogram method was used to extract CECT texture features of thymic epithelial tumors. The mean, variance, skewness, and kurtosis based on the gray-level histogram were measured to describe the dispersion, distribution characteristics, and gray peak of the pixel distribution.⁹ Energy is a measure of the uniformity of the image gray-scale distribution, and entropy describes the complexity of the gray distribution in a given region of interest.⁹ The results showed significant differences in mean, skewness, and entropy between low-risk and high-risk thymic epithelial tumors, with entropy having the best diagnostic value. The study findings suggest that CECT texture features can be used to identify high-risk thymic epithelial tumors.

Aggressiveness and heterogeneity of tumors are closely related.²⁹ Computed tomography texture analysis is a new image postprocessing technique that can be used as a tool to quantify the heterogeneity of tumors.³⁰ Previous authors have used CT texture features to differentiate high-risk from low-risk thymic epithelial tumors.^25–28 Sui et al²³ reported that the mean in unenhanced CT, and skewness and kurtosis in CECT models, showed good performances in distinguishing high-risk tumors from low-risk tumors in the anterior mediastinum. Iannarelli et al²⁶ demonstrated a significant relationship between texture features (mean, skewness, kurtosis, entropy) and histology of thymic tumors. Rajamohan et al²⁸ used a filtration histogram to generate 6 first-order texture parameters (mean, SD, mean of positive pixels, entropy, skewness, and kurtosis) and found that none of the parameters were significantly different between WHO low-grade and high-grade tumors. Our findings are consistent with the results reported by Iannarelli et al²⁶ and Sui et al.²³ Among the 6 first-order texture parameters, entropy showed the best diagnostic performance. Entropy >7.74 suggests a greater likelihood of malignant thymic epithelial tumor. Differences in pathological classification, clinical stage of study patients, CT scanning parameters,³¹ and texture analysis algorithm³² may be responsible for the different results reported by Rajamohan et al.²⁸

Computed tomography texture analysis has also shown promise for hepatocellular carcinoma, gastric cancer, head and neck cancer, and renal tumors.^11,13,15,17 Raman et al¹¹ used a random forest model and showed that clustered CT texture features of mean and entropy values could reliably distinguish cysts from oncocytoma, clear-cell renal cell carcinoma, and papillary renal cell carcinoma. Oh et al¹³ reported that CT texture feature of skewness on arterial-phase images was an independent predictor of disease-free survival in patients with hepatocellular carcinoma. Giganti et al¹⁵ showed that preoperative CT texture features of entropy and skewness were significantly associated with poor prognosis in gastric cancer. Zhang et al¹⁷ found that the entropy and skewness of primary head and neck squamous cell carcinoma, measured with the 1.0 spatial filter, were independently associated with overall survival. These findings are consistent with our study that showed that entropy, mean value, skewness, and kurtosis can be used to distinguish high-risk from low-risk thymic epithelial tumor, with entropy having the best diagnostic value.

The present study has several limitations. First, this was a single-center retrospective study with a small sample size, so bias is inevitable. Second, we did not compare the diagnostic performance of the gray-level histogram texture algorithm with that of other algorithms. Third, we compared texture differences only on CECT images; we did not assess unenhanced images although Sui et al²³ found that texture analysis of unenhanced CT images has higher diagnostic efficiency for distinguishing high-risk anterior mediastinal lesions. Finally, this work is based on a single scanner, single reconstruction method, and single kernel; it is possible that scanner model, reconstruction method, and other variables influenced our results.

In conclusion, CECT texture features appear to have value for identifying high-risk thymic epithelial tumors. The texture analysis method could be a useful clinical tool to guide treatment decisions.