Preoperative risk stratification of thymic epithelial tumors using multimodal CT habitat radiomics and Vision-Mamba deep learning: a multicenter retrospective study

Preoperative distinction between low- and high-risk thymic epithelial tumors is difficult because biopsy may not capture intratumoral heterogeneity. This study developed and evaluated a multimodal computed tomography model combining habitat radiomics, differential elasticity mapping, and Vision-Mamba for preoperative risk stratification. This multicenter retrospective study included 337 patients with pathologically confirmed thymic epithelial tumors from two centers between 2015 and 2025. Patients from one center were randomly assigned to a training cohort ( n = 208) and an internal validation cohort ( n = 90); patients from the second center formed an independent external test cohort ( n = 39). Tumors were grouped as low risk (types A, AB, and B1) or high risk (types B2, B3, and thymic carcinoma) according to the 2021 World Health Organization classification. Models were developed using venous-phase contrast-enhanced computed tomography images, computed tomography-derived differential elasticity maps, and habitat radiomics features. Performance was assessed using the area under the receiver operating characteristic curve, calibration curves, the Hosmer–Lemeshow test, Brier scores, decision curve analysis, and DeLong tests. A conventional radiological model was used as a comparator. The combined model achieved areas under the curve of 0.948 (95% confidence interval, 0.921–0.971), 0.937 (0.887–0.975), and 0.902 (0.796–0.976) in the training, internal validation, and external test cohorts, respectively. Corresponding accuracies were 0.851, 0.856, and 0.846, and Brier scores were 0.120, 0.131, and 0.141. In the external test cohort, differences between the combined and individual imaging models were not statistically significant. The combined model outperformed the conventional radiological model in all three cohorts; external-test areas under the curve were 0.902 versus 0.612 ( P = 0.007). Decision curve analysis indicated greater net benefit for the combined model across most threshold probabilities. The multimodal computed tomography model showed promising performance for preoperative risk stratification of thymic epithelial tumors and provided incremental value over conventional radiological assessment. Because the external cohort was small, larger prospective multicenter studies are required before clinical implementation.