Multi-modal prediction model for thymic epithelial tumors: Enhancing surgical decisions and recurrence risk assessment from CT datasets.

Background: Precise prediction of histology (TETs vs. Non-TETs), Masaoka-Koga staging, and WHO classification is essential for surgical and postoperative decisions. Artificial intelligence (AI) models have shown promise in addressing these gaps, but existing studies often suffer from small datasets and single-task limitations. Purpose: This study aims to develop and validate a multi-task deep learning model for preoperative non-invasive prediction of histology and Masaoka-Koga stage in thymic epithelial tumors to support individualized risk assessment and treatment planning. Materials and Methods: We analyzed 659 chest CT scans from TET patients with pathologically confirmed histology, Masaoka-Koga staging, and WHO classification collected between October 2014 and July 2025 from Shanghai General Hospital, Shanghai Pulmonary Hospital, and Huashan Hospital. A deep learning model was developed to classify histology into TETs and Non-TETs groups, Masaoka-Koga stages and WHO risk groups. Results: The model achieved high performance across tasks in the test set. For histology classification, AUC was 0.9674 (95% CI, 0.9347-0.9895), with accuracy of 91.73% (95% CI, 86.47%-96.24%). Masaoka-Koga staging stratification yielded an AUC of 0.9328 (95% CI, 0.8179-1.0000) and accuracy of 95.38% (95% CI, 89.23%-100.00%). WHO risk prediction reached an AUC of 0.8485 (95% CI, 0.7447-0.9347) and accuracy of 78.46% (95% CI, 67.69%-89.23%). Calibration analysis further showed low Brier scores for the histology and Masaoka-Koga models (0.0695 and 0.0456, respectively), supporting acceptable probability calibration for these two tasks. Conclusion: This multi-task AI model provides a non-invasive, accurate tool for preoperative TET assessment, potentially optimizing surgical strategies and personalized care.