Abstract
Accurate preoperative risk stratification is critical for treating thymic epithelial tumors (TETs). This study developed a deep learning framework that combines a dual-channel convolutional neural network (CNN) with an adaptive dynamic clustering algorithm. The model was trained on contrast-enhanced CT (CECT) images from 336 multicenter TET patients. It first automates the segmentation of tumor subregions. Then, it constructs dual-channel input data containing the largest cross-sectional ROI and its corresponding habitat masks. Using transfer learning, we trained four CNN architectures for risk stratification. The DenseNet121-based dual-channel CNN achieved an AUC of 0.74-0.76 on an external test set. This performance surpassed conventional radiomics, single-channel CNN, and radiologists' visual assessment. Our framework effectively captures intratumoral heterogeneity, improves risk stratification accuracy, and aids in rapid identification of high-risk patients. This approach can support personalized treatment planning.