Machine learning-based radiomics for bladder cancer staging: evaluating the role of imaging timing in differentiating T2 from T3 disease

Objectives Accurate preoperative staging of bladder cancer is essential for therapeutic decision-making, particularly in distinguishing between organ-confined (T2) and extravesical (T3) disease. This study aimed to develop a CT-based radiomics model to differentiate T2 from T3 tumors and to evaluate the impact of imaging timing relative to transurethral resection of the bladder (TURB) on model performance. Additionally, we assessed the added diagnostic value of integrating routine clinical biomarkers. Methods In this retrospective study, 97 patients with histologically confirmed bladder cancer who underwent TURB followed by contrast-enhanced CT were included. Tumor segmentation was performed using a semi-automated three-dimensional approach, and radiomic features were extracted according to IBSI standards. A random forest classifier was trained to distinguish between T2 and T3 tumors. Patients were stratified according to the interval between TURB and CT imaging (≤14 days vs 14 days). Performance metrics were assessed for both radiomics-only and combined clinical-radiomics models. Clinical variables included preoperative creatinine, hemoglobin, arterial hypertension, diabetes mellitus, smoking status, and tumor size. Results The radiomics-only model achieved an AUC of 0.68 in Cohort 1 (≤14 days post-TURB). In Cohort 2 (14 days post-TURB), model performance improved with an AUC of 0.80. The combined clinical-radiomics model further enhanced performance, yielding an AUC of 0.76 in Cohort 1 and 0.82 in Cohort 2. Delayed imaging was associated with increased radiomic feature stability and improved classification accuracy, suggesting a potential benefit of temporal separation from post-surgical tissue changes. Conclusion This study demonstrates the feasibility of CT-based radiomics using full-volume 3D tumor segmentation to distinguish between T2 and T3 bladder cancer. The integration of clinical biomarkers and consideration of imaging timing significantly improved model performance. These findings support the development of temporally optimized, multimodal prediction models for individualized bladder cancer staging and treatment planning.