Risk classification of thymoma based on multi-feature fusion in dynamic enhanced CT

Area of Science:

• Oncology
• Radiology
• Artificial Intelligence in Medicine

Background:

• Thymoma classification into high-risk and low-risk groups is crucial for treatment planning and prognosis.
• Current methods face limitations in fully utilizing imaging data, especially combining radiomics and deep learning for preoperative classification.

Purpose of the Study:

• To develop and validate a comprehensive computed tomography (CT) imaging model (CSRT) for enhanced preoperative classification of thymomas.
• To evaluate the model's utility in non-invasive diagnosis and risk stratification.

Main Methods:

• A retrospective study of 360 thymoma patients from three centers was conducted.
• CT images (NECT, CECT) were used to extract radiomics and Vision Transformer (ViT)-based deep learning features.
• Clinical semantic features were integrated, and key features were selected using t-tests and LASSO regression to build the fusion model.

Main Results:

• The CSRT model achieved an AUC of 0.835, accuracy of 77.9%, sensitivity of 78.4%, and specificity of 77.1% in the external validation cohort.
• Calibration curves demonstrated high consistency between predicted and actual classifications.
• Decision curve analysis confirmed clinical utility, showing high net benefit above a 30% threshold probability.

Conclusions:

• The CSRT model effectively differentiates high-risk and low-risk thymomas preoperatively using CT data.
• This non-invasive tool supports individualized treatment strategies and enhances clinical decision-making in thymoma management.
• The model provides reliable classification above clinically relevant risk thresholds.