Development and validation of a CT-TA nomogram for preoperatively differentiating thymic epithelial tumor histologic subtypes

Background: Thymic epithelial tumors (TETs) are the most common primary tumors in the anterior mediastinum with considerable histologic heterogeneity. This study aimed to develop and validate a nomogram based on computed tomography (CT) and texture analysis (TA) for preoperatively predicting pathological classification for TET patients. Methods: Totally 172 patients with pathologically confirmed TET after surgery were retrospectively analyzed and randomly divided into a training cohort (n=120) and a validation cohort (n=52). Preoperative clinical demographic, CT signs and texture features of each patient were analyzed and prediction models were developed using the least absolute shrinkage and selection operator (LASSO) regression. The performance of models was evaluated and compared by the area under receiver operating characteristic (ROC) curve (AUC) and DeLong test. The clinical application value of models was determined through the decision curve analysis (DCA). Then a nomogram was developed based on the model with the best predictive accuracy and clinical utility and validated using the calibration plots. Results: Totally 87 patients with low-risk TET (LTET) (types A, AB, B1) and 85 patients with high-risk TET (HTET) (types B2, B3, C) were enrolled in this study. We separately constructed 4 prediction models for differentiating LTET from HTET using clinical, CT, texture features and their combination. These 4 prediction models achieved AUC values of 0.66, 0.79, 0.82, 0.88 in the training cohort and 0.64, 0.82, 0.86, 0.94 in the validation cohort, respectively. DeLong test and DCA analysis showed that the Combined model consisting of 2 CT signs and 2 texture parameters held the highest predictive efficiency and clinical utility ( p <0.05). A prediction nomogram was subsequently developed using the combined model’s 4 independently risk factors. The calibration curves indicated a good consistency between the actual observation and nomogram prediction for

Conclusion: A prediction nomogram incorporating both the CT and texture parameters was constructed and validated in our study, which was conveniently used to facilitate the preoperative individualized prediction of simplified histologic subtypes in TET patients, assisting in clinical treatment decision making and achieving precision treatment.

Background
Thymic epithelial tumors (TETs), which are the most common primary tumors in the anterior mediastinum, are well-known for heterogeneity in the oncologic and biologic behaviors [1]. According to the morphology of epithelial cells as well as the lymphocyteto-epithelial cell ratio, the World Health Organization (WHO) classification which was proposed in 2015 classified TETs into six subtypes (thymoma: types A, AB, B1, B2 and B3; thymic carcinoma: type C), which is recognized as the basis of the clinical treatment decision making and an independent prognostic factor in TETs [2][3][4]. Previous studies have shown that the invasiveness of each subtype increased in turn, and patients in low-risk TET (LTET) (types A, AB, B1) usually had more chances to be completely resected, lower tumor recurrence rate and higher survival rate than ones in high-risk TET (HTET) (types B2, B3, C) [5][6][7]. Moreover, HTET patients can benefit more from adjuvant treatment than LTET patients [8]. Thus, preoperative knowledge of histologic classification can provide valuable information for treatment decision making and prognostic evaluation in TET patients.
Currently, at least 3 imaging modalities are used for TET patients. Although positron emission tomography-computed tomography (PET/CT) is the most sensitive and accurate method for tumors [9], it is not recommended as a routine examination because of the high cost and great radiation damage [10]. Magnetic resonance imaging (MRI) may be another method that could provide functional quantitative indicators for TETs [11], but it is not appropriate for patients with mental implants or poor respiratory control. Computed tomography (CT) is widely recognized as the main imaging method for the diagnosis, differentiation and evaluation of curative effect in TET patients due to its convenient operation, good image quality, moderate price and less contraindications [12]. The signs on CT images, such as size, location of the lesions, as well as the presence of pericardium or pleural effusion and distant metastases, are helpful to preliminarily judge the invasiveness of TETs [13][14][15]. However, they are of limited value for further accurate assessment tumor heterogeneity or differentiation of its histological subtypes [16].
Texture analysis (TA) based on conventional medical images has been used to quantity assessment of tumor heterogeneity by analyzing the distribution and relationship of pixel or voxel gray levels in the lesion area [17; 18]. Previous studies have revealed that TA, as a non-invasive imaging tool, has great potential in predicting pathologic features, response to therapy and prognosis of head and neck cancer, rectal cancer, et.al [19; 20].
Moreover, CT quantitative TA has been conducted for preoperative assessment of TETs and shown good diagnostic performance in differentiating tumor types and stages [21; 22]. However, only texture parameters extracted from CT images were analyzed in above studies, whether the prediction model combines with CT signs and texture parameters can further improve the discrimination performance is an interesting problem that requires investigation.
Hence, the aim of this study was to develop and validate a nomogram consisting of CT morphological features and texture parameters for differentiating the simplified histologic subtypes in patients with TETs.

Patients
We conducted a retrospective analysis of patient records with TETs diagnosed between January 2011 to April 2019 at two cancer centers: Shanghai Proton and Heavy Ion Center (institution A) and Fudan University Shanghai Cancer Center (institution B). This study was approved by the ethics committees of these two institutions and the requirement for informed consent was waived. The inclusion criteria including the following: 1) underwent standard contrast-enhanced CT at above two institutions less than 14 days before surgery; 2) patients who underwent surgery at institution B for TETs with curative intent; 3) information of postoperative pathologically confirmed TETs histologic subtypes available.
The exclusion criteria including the following: 1) previous history of TETs or other cancer; 2) preoperative therapy (radiotherapy, chemotherapy or chemoradiotherapy); 3) poor image quality affects lesion segmentation. Patients were divided into a training cohort (n=120) and a validation cohort (n=52) after simple randomization at a ratio of 7 to 3. Baseline data pertaining to demographics of each patient, including gender, age, symptom was reviewed and recorded.

CT images acquisition and analysis
Patients generally underwent contrast-enhanced CT of the entire thorax according to

Tumor segmentation and texture feature extraction
Tumor segmentation and feature extraction were performed using the LIFEx (version 5.10, CEA-SHFJ, Orsay, France) package. The above two radiologists selected the largest slice of the tumor at three-dimensional (3D) images to delineate the region of interest (ROI) by consensus ( Fig. 1 a-d). The ROI selection should include all tumors and avoid blood vessels, calcification and gas.
A total of 43 texture features were extracted automatically [23], including 2 shape features. The details of texture features were described in supplementary data.

Statistical analysis
Statistical analysis was performed in R (version 3.6.0, http://www. r-project. org). A twotailed p value of < 0.05 was used as the criterion to indicate a statistically significant difference. The Mann-Whitney U test was used to assess differences in continuous variables, whereas the χ2 test was used for categorical variables. Intra-and interclass correlation coefficients (ICCs) were used to evaluate the consistency and reproducibility of the intra-and inter-observer agreement of extracting features by two-level radiologists.
An ICC greater than 0.75 indicated good consistency.

Prediction performance and Clinical utility of predictive models
The performance of the models was evaluated by the area under the receiver operating characteristic (ROC) curve (AUC) and compared by Delong test. The AUCs, sensitivity, specificity and accuracy with 95 % confidence intervals (CIs) were calculated for each model. The clinical application value of the prediction models was determined through the decision curve analysis (DCA) by quantifying the net benefit to the patient under different threshold probabilities in the queue.

Development and validation of nomogram
To provide a visually quantitative tool to predict histologic subtype for TET patients, we develop a nomogram on the basis of the prediction model with the highest AUC value and clinical utility on training cohort. Calibration curve was plotted to assess the calibration of the nomogram by bootstrapping (1000 bootstrap resamples) based on internal (training cohort) and external (validation cohort) validity.

Feature selection and prediction model establishment
A total of 12 CT signs and 43 texture features were extracted from 172 TET patients' enhanced CT images, and the agreement between the two radiologists (readers 1, 2) was excellent for texture features (all ICCs >0.85, p<0.05). Thus, the mean measurement values of the two radiologists were used for further analysis.
The cross-correlation matrixes showed that there were multiple complex cross-correlations among these parameters (Fig. 2). For differentiating LTET from HTET, 4 independent prediction models were built separately on the basis of selected clinical, CT, texture parameters and their combination by LASSO regression in the training cohort (Fig. 3 a-d).
The Pre-scores of each model for each patient were calculated using the following HTET patients generally had higher Pre-scores for all prediction models than LTET patients in both the training and validation cohorts (p <0.05, respectively) ( Table 2).

Prediction performance and Clinical utility of predictive models
The performance of these 4 models to discriminate LTET from HTET is shown in Table 3,  Table 3).
The DCAs also showed that the clinical application value of these 4 prediction models increased in turn, which indicated that the Combined model incorporating CT morphological features and texture parameters was a most reliable clinical treatment tool for predicting histologic subtypes in TET patients when the threshold probability was between 0.02 and 0.91 (Fig. 5).

Development and validation of nomogram
According to the above results, we generated a nomogram based on the parameters of the Combined model for visualization (Fig. 6). The calibration curves of nomogram for the probability of HTET demonstrated a good agreement between prediction by nomogram and actual observation in two cohorts (p >0.05, respectively) (Fig. 7).

Discussion
In the present study, we developed 4 independent prediction models to differentiate the pathologic subtypes of TETs, and constructed a quantitative nomogram based on the model which held the highest efficiency and clinical utility. This nomogram was validated for the preoperative individualized prediction of the classification in TET patients.
In terms of clinical characteristics and CT signs, although the correlation between patient's gender, age and tumor invasiveness is still controversial [2; 4; 15], our study indicated that an older female TET patient with a bigger tumor size, a smoother boundary and a clearer mediastinum fat line indicates a lower probability that the tumor is malignant. In this study, tumor size, boundary and mediastinum fat line were significantly associated with the malignant grade of TETs. According to the general understanding, the larger the tumor, the more malignant it is [27]. However, the size in LTET was larger than that in HTET in both the training and validation cohorts (p values < 0.05, respectively). This may be related to the fact that the less aggressive the tumor is, the lighter clinical symptoms the patient has, which leads to a larger volume when it is found. The boundary and mediastinum fat line of tumor on CT images can reflect the invasiveness of the mass to a certain extent [28; 29]. There is usually a higher rate of tumor invasiveness by directly extending to adjacent structures including vessels, pericardium or lung, which is shown as a rough boundary and blurred or even disappeared fat line. The results of this study are basically consistent with previous researches.
In this study, the texture parameters of meanValue, sphericity and NGLDM_Busyness of LTET patients were higher than HTET ones. Meanwhile, the TA model composed of these 3 parameters held great individualized prediction for TET patients (AUCs = 0.82 [training cohort], 0.86 [validation cohort], respectively). MeanValue in histogram, which represent the average value of ROI, reflects the degree of texture regularity: the higher the value, the more regular the texture is, that is, the lower the heterogeneity is. Yasaka K et al. also found that the meanValue was a significant parameter for differentiating between HTET and LTET with AUC of 0.89 [21]. Sphericity is the shape features of tumors, and has been proven as the most significant affecting factor for discriminating histologic subtype in TET patients [30]. Busyness is a parameter of NGLDM, which measures the spatial frequency of changes in intensity between nearby voxels of different grey-levels. The role of busyness in TETs has not been reported before, but it has been used to assess the tissue However, the present study had several limitations although the results were encouraging.
First, the sample selection was biased in this retrospective study, and a prospective study is required to confirm and validate the prediction nomogram. Second, the texture features extracted in this study were only based on enhanced CT images, but not the plain CT.
However, previous study showed that there was no significant difference between radiomics features based on plain CT images and ones based on enhanced CT images for predicting risk categorization of TETs [30]. Third, the tumor, node, metastasis (TNM) [39] or Masaoka [40] staging systems for TETs were not used in this study. Further study will be needed to reveal the relationship with texture features and TNM or Masaoka staging systems.

Conclusions
A prediction nomogram incorporating both the CT morphological features and texture parameters was constructed and validated in our study, which was conveniently used to

Availability of data and materials: Yes.
Competing interests: The authors declare that they have no competing interests.     Note: AUC, area under the receiver operating curve; 95% CI: 95% confidence interval. Figure 2 The cross-correlation matrix for covariates. Blue represents positive correlation and red represents negative correlation. The depth of color indicates the intensity of the correlation between covariates. The darker the color, the higher the correlation is.   DCA for the prediction models. The X-axis represents the threshold probability.

Figures
The Y-axis represents the net benefit. The grey and black thin lines represent the hypothesis that all TET patients are high-risk or low-risk, respectively. The higher curve at any given threshold probability is the optimal prediction to maximize net benefit.