A CT-based deep learning radiomics model for predicting HER2 expression and prognosis in non-muscle-invasive bladder cancer.

OBJECTIVE

This study aimed to extract radiomics (Rad) and deep learning (DL) features from preoperative CT of patients with Non-Muscle Invasive Bladder Cancer (NMIBC) and develop models incorporating clinical characteristics to assess Human-Epidermal-Growth-Factor-Receptor-2 (HER2) expression status and prognosis in these patients.

METHODS

From January 2019 to December 2024, 181 patients with NMIBC were retrospectively enrolled in this study. A deep learning radio-clinical signature model (DLCS) was created by integrating DL score, Rad score, and clinicopathologic features to predict HER2 expression in NMIBC and compared with a deep learning model, a radiomic model, and a Clinical model. An additional model was built to predict Recurrence-Free Survival (RFS) in NMIBC patients.

RESULTS

181 patients with NMIBC were divided into a training cohort (n = 126) and a test cohort (n = 55). The DLCS model achieved the highest area under the curve (AUC) for HER2 prediction in the test cohort (AUC = 0.894 (95% CI: 0.814-0.974)). The univariate and multivariate Cox regression analyses identified both the DL score and Rad score as independent risk factors for RFS (p < 0.05).

CONCLUSION

The DLCS model demonstrates good diagnostic performance in predicting HER2 expression, and the prognosis model can stratify the risk of tumor recurrence in patients with NMIBC.