Prediction of pathologic upstaging in ductal carcinoma in situ using MRI radiomics combined with clinicopathological features.

BACKGROUND: Core needle biopsy (CNB) for ductal carcinoma in situ (DCIS) diagnosis is limited by sampling error, potentially leading to under-staging and suboptimal management. This study aimed to develop a predictive model for DCIS upstaging based on preoperative MRI radiomics and clinicopathologic features, to assist in preoperative risk stratification and clinical decision-making. Additionally, we explored the potential value of intratumoral heterogeneity (ITH) as an adjunctive imaging biomarker. METHODS: Preoperative MRI images and clinicopathological data were retrospectively collected from patients diagnosed with DCIS by CNB and treated surgically at our institution between December 2018 and April 2023. Radiomics features were extracted from tumor volume of interest and used to generate a radiomics score via LASSO regression. Tumor subregions were clustered by K-means based on signal intensity to derive an ITH score. Univariate and multivariate logistic regression analyses were performed to identify independent predictors of DCIS upstaging. Several multivariate logistic regression models were then developed based on these predictors. Model performance was assessed using the area under the receiver operating characteristic curve (AUC), calibration, and decision curve analysis. The final model was visualized with a nomogram. RESULTS: A total of 237 DCIS lesions from 232 patients (median age, 49 years [IQR, 43, 55]) were analyzed. Univariate analysis identified several factors associated with upstaging, including Ki67 ≥ 14%, higher nuclear grade, higher biopsy needle gauge, larger lesion size, and higher radiomics and ITH scores (all P < 0.05). In multivariate logistic regression analysis, nuclear grade (OR range, 5.19–6.61[95% CI 1.0, 47.8]; all P < 0.05), biopsy needle gauge (OR, 1.17 [95% CI 1.02, 1.38]; P = 0.039), and radiomics score (OR, 1.95 [95% CI 1.25, 3.17]; P = 0.004) remained independent predictors. However, the ITH score was not independently predictive in the multivariate model and were excluded from the final model. The combined clinicopathologic and radiomics model demonstrated superior predictive performance with an AUC of 0.82 (95% CI 0.70, 0.93) in the test set, outperforming both the clinicopathologic-only (AUC = 0.73) and radiomics-only (AUC = 0.75) models. CONCLUSIONS: The combination of MRI radiomics score and clinicopathologic features effectively predicted DCIS upstaging, supporting their value in preoperative invasive risk assessment and treatment decision-making.