Multicenter development and external validation of clinical-radiomics models to predict surgically confirmed
Xi Hu1, Lujie Qian1, Jie He1
1Department of Radiology, Sir Run Run Shaw Hospital, Zhejiang University School of Medicine, Hangzhou, China.
Purpose:
Ductal carcinoma in situ (DCIS) diagnosed on core biopsy is frequently upgraded to invasive carcinoma at surgery, which may change indications for sentinel lymph node biopsy. Routine breast MRI has limited ability to detect occult invasion preoperatively. This study aimed to develop and externally validate an MRI-based model combining clinical variables, conventional MRI findings, and dynamic contrast-enhanced (DCE) MRI radiomics to predict invasive upgrade in biopsy-proven DCIS.
Methods:
This retrospective multicenter study enrolled 478 patients from three hospitals (2014-2019). Center 1 contributed 314 patients, randomly split into a training set (n = 251) and an internal test set (n = 63); Centers 2 (n = 39) and 3 (n = 62) formed two independent external test sets. Radiologists assessed conventional MRI features, including lesion size, enhancement descriptors, and diffusion-derived apparent diffusion coefficient metrics. Tumors were segmented on DCE MRI. Radiomics features with intraclass correlation coefficient > 0.85 were z-score normalized, selected using least absolute shrinkage and selection operator regression, and used to train multiple machine learning classifiers; the best-performing model generated a radiomics score. Model selection and hyperparameter tuning were performed by cross-validation within the training set only. Clinico-radiologic, radiomics, and combined models were evaluated using receiver operating characteristic (ROC) curve analysis, calibration, and decision curve analysis, the area under the curve (AUC) was calculated.
Results:
Six clinico-radiologic factors and 13 radiomic features were retained. In the two external test sets, the clinico-radiologic, radiomics, and combined models achieved AUCs of 0.61 (95% CI, 0.43-0.79) and 0.71 (0.58-0.83), 0.70 (0.54-0.86) and 0.71 (0.58-0.84), and 0.76 (0.60-0.91) and 0.77 (0.65-0.89), respectively. The combined model provided the highest net benefit on decision curve analysis.
Conclusion:
A combined clinico-radiologic and DCE-MRI radiomics model showed multicenter, externally validated performance for preoperative prediction of invasive upgrade in DCIS, supporting risk stratification for surgical planning.
