Introduction

Thymic epithelial tumors (TETs) represent a heterogeneous group of rare tumors, but they are the most common anterior mediastinal tumors that have low to high malignant potential ¹. It is critically important to accurately predict the pathological subtype and TNM staging for TET patients. Imaging plays an important role in the preoperative diagnosis, staging, and follow-up monitoring of TETs ². Recently, radiomics has been used to predict tumor structural, pathologic, and genetic characteristics based on quantitative features extracted from medical images ³. In this study, we aimed to develop predictive models based on MR radiomics features for identifying the pathological classification and TNM staging of TETs, and to generate a radiomics nomogram that incorporated the radiomics model, MR imaging findings and clinical variables for predicting the risk of advanced TETs.

Methods

This retrospective single-center study was approved by the local Ethics Committee, and informed consent was waived. Between November 2014 and November 2017, a total of 189 consecutive TET patients with MR imaging evaluation of the thorax were enrolled in this study. All patients were randomly assigned in a 3:1 ratio to either the training cohort or test cohort. Conventional MR imaging findings (peritumoral edema, pleural effusion, pericardial effusion, invasion of adjacent tissues) were assessed by two senior radiologists. ITK-SNAP software was used for three-dimensional manual segmentation on all axial T2W and FS images. Using a noncommercial Analysis-Kit software, the method of quantitative feature extraction was conducted on T2W and FS images. The quantitative features from each sequence were independently re-ranked using the Support Vector Machine Recursive Feature Elimination (SVM-RFE) method. The top-ranked quantitative features subsets with the highest accuracy from T2W and FS images would be combined for feature selection. Machine learning-based radiomics models were built based on the established optimal feature subsets and SVM classifier. The classification accuracy, sensitivity, specificity and area under the curve (AUC) were measured in both the training and test cohorts. To predict the probability of advanced TETs, multivariable logistic regression analysis was performed with the radiomics model, clinical and imaging variables.

Results

The radiomics model for pathological classification using the combination of T2W and FS images achieved better discriminative ability than those using either of them alone, with the accuracy, sensitivity, specificity, and AUC of 75.9%, 74.8%, 87.9% and 88.0%, respectively. The model was then applied to the test cohort, and the accuracy, sensitivity, specificity, and AUC were 68.9%, 67.5%, 84.4% and 77.1%, respectively. The radiomics model for TNM staging using the combination of T2W and FS images also achieved better discriminative ability than those using either of them alone, with 92.6% accuracy, 92.7% sensitivity, 90.5% specificity, and 92.8% AUC in the training cohort, and 89.4% accuracy, 91.7% sensitivity, 87.5% specificity, and 90.8% AUC in the test cohort. According to the multivariable logistic regression analysis, pericardial effusion (P = 0.026) and radiomics model (P < 0.001) represented independent predictive factors of advanced TETs. Despite lack of independent predictive status, the presence of symptom was left in the final nomogram as it increased predictive accuracy (Figure 1). The C-index in the training and test cohorts were 0.937 (95% confidence interval [CI]: 0.857, 0.955) and 0.957 (95% CI: 0.842, 0.974), respectively. Figure 2 and 3 show how our radiomics nomogram could be used to calculate the predicted probability of advanced TETs.

Discussion

In this study, we developed two predictive models based on MR radiomics features to differentiate the pathological subtype and TNM stage of TETs, which achieved high predictive efficacy. Additionally, the radiomics nomogram yielded good performance and could facilitate the preoperative individualized assessment of TETs. We found that the radiomics model using the combination of T2W and FS images achieved higher predictive ability than those using either of them alone. It is not surprising that biparametric MR images can provide more information than single ones. Compared with the radiomics model, the nomogram revealed improved predictive accuracy after the multivariable model modification. This finding may support the idea that taking into account markers that span different aspects is the most promising approach to change clinical management. Several limitations to this study should be considered. First, radiomics analysis was performed using a single machine in a single institution with same MRI protocol on mass lesions only. Second, we only used T2W and FS images for radiomics analysis, which reduced variability in acquisition parameters.

Conclusions

Our preliminary results reveal that quantitative features extracted from MR images are closely related to the biological behavior of TETs, and radiomics models could facilitate the accurate prediction of pathological classification and TNM staging in TETs.

Acknowledgements

No acknowledgement found.