Abstract Background Perioperative risk stratification in thymoma-associated myasthenia gravis remains challenging, and the optimal spatial extent for preoperative computed tomography (CT) radiomics is uncertain. We compared five three-dimensional tumor and peritumoral radiomic representations for discrimination of an in-hospital perioperative composite outcome. Methods This retrospective single-center study included 145 patients (37 events). Intratumoral, physical-distance peritumoral (1, 3, and 5 mm), and integrated tumor-plus-3-mm (VOIFusion3mm) representations were compared. Radiomic models used radial-basis-function support vector machines with five-fold nested cross-validation repeated 10 times. Prespecified paired comparisons used 2,000 patient-cluster bootstrap resamples with false-discovery-rate correction. Feature-exclusion, no-LASSO, boundary-perturbation, and clinical-increment analyses were also performed. Results The integrated tumor-plus-3-mm representation had the highest mean repeat-level out-of-fold area under the receiver operating characteristic curve (AUC) (0.628; 95% confidence interval (CI), 0.544–0.714), versus 0.536 for the intratumoral and 0.533 for the 3-mm peritumoral representation. AUC differences were 0.092 versus intratumoral (95% CI, 0.029–0.155; q = 0.009) and 0.095 versus 3-mm peritumoral (95% CI, 0.024–0.173; q = 0.015). VOIFusion3mm remained highest-ranked after feature exclusion and in the no-LASSO analysis. Mean AUC changed by − 0.093 after 1-mm inward erosion (95% CI, − 0.169 to − 0.015) and by − 0.004 after outward dilation (95% CI, − 0.028 to 0.019). Radiomics did not demonstrate incremental discrimination beyond the prespecified clinical model, and calibration was limited. Conclusions VOIFusion3mm showed higher internally evaluated rank discrimination than the intratumoral and isolated 3-mm peritumoral representations, but absolute discrimination was modest. Radiomics did not demonstrate incremental discrimination beyond the evaluated clinical models, and calibration did not support individualized risk estimation. VOIFusion3mm remains a candidate spatial representation requiring external validation and should not guide individual perioperative decisions. Interpretation is limited by the small number of outcome events, single-center design, and lack of independent multi-reader assessment of segmentation reproducibility.