Evaluating the incremental value of radiomics and pretrained breast MRI foundation-model representations beyond conventional preoperative features for predicting invasive upgrade in biopsy-proven ductal carcinoma in situ
Abstract Background Preoperative identification of invasive carcinoma in patients diagnosed with pure ductal carcinoma in situ (DCIS) on percutaneous biopsy remains challenging because of sampling-related underestimation. We evaluated whether multiparametric magnetic resonance imaging (MRI) radiomics and pretrained breast MRI foundation-model representations provide incremental discrimination beyond conventional preoperative information for invasive upgrade. Methods This retrospective single-center study included 166 DCIS lesions in 164 patients who underwent preoperative breast MRI and surgical excision; 66 lesions (39.8%) were upgraded to invasive carcinoma. Predictors comprised conventional preoperative demographic, biopsy pathology, MRI, and mammographic variables; multiparametric MRI radiomics; and frozen representations from two pretrained breast MRI foundation models, Pillar-0 and FM-BCMRI. A patient-level 30% held-out test set was fixed before model development. Fifteen feature-block combinations were evaluated using patient-grouped five-fold cross-validation in the development set. The final model was selected according to the development out-of-fold (OOF) area under the receiver operating characteristic curve (AUC). Incremental AUCs were assessed using paired DeLong tests with Benjamini–Hochberg adjustment. Calibration, decision-curve analysis, and SHAP-based interpretability were also evaluated. Results The conventional preoperative + Pillar-0 model achieved the highest development OOF AUC (0.75; 95% confidence interval (CI), 0.64–0.83), versus 0.66 (95% CI, 0.55–0.76) for the conventional preoperative model. In the held-out test set, the final model achieved an AUC of 0.65 (95% CI, 0.49–0.79), whereas the conventional model achieved 0.74 (95% CI, 0.58–0.85). Adding Pillar-0 yielded a development OOF ΔAUC of + 0.085 ( P = .14; adjusted P = .48) and a held-out ΔAUC of − 0.086 ( P = .33; adjusted P = .41). No imaging-augmented conventional model showed statistically significant incremental discrimination after multiple-comparison adjustment. Held-out calibration and decision-curve analyses showed no consistent advantage of the final model over the conventional comparator. Conclusions Conventional preoperative information showed moderate discrimination for invasive upgrade in biopsy-proven DCIS. Although radiomic and foundation-model representations showed apparent discrimination during development, their addition did not provide statistically supported or consistent incremental benefit in held-out evaluation. Larger independent multicenter cohorts are needed to determine whether these imaging-derived representations provide reproducible complementary information for preoperative DCIS risk stratification.
Authors
- Fatma Çelik Yabul (ORCID: https://orcid.org/0000-0002-0156-0056)
- Süleyman Öncü (ORCID: https://orcid.org/0000-0002-1434-3536)
- Güler Taşdemir (ORCID: https://orcid.org/0009-0007-9597-0008)
- Murat Yüce (ORCID: https://orcid.org/0000-0003-0619-5868)
- Vesile Kaya (ORCID: https://orcid.org/0009-0007-5683-744X)
Institutions
- Bakırköy Dr.Sadi Konuk Eğitim ve Araştırma Hastanesi (TR)
- Icahn School of Medicine at Mount Sinai (US)
Publication Details
- Journal
- BMC Medical Imaging
- Published
- 2026-11-17
- DOI
- https://doi.org/10.1186/s12880-026-03067-8
- Primary Topic
- Radiomics and Machine Learning in Medical Imaging
- Type
- article
- Field-Weighted Citation Impact
- 0.00