Automated Cluster-based Quantitative Analysis of Ultrafast DCE MRI for Differential Breast DCIS Grading

Purpose To assess whether k-means clustering (KMC) analysis of ultrafast dynamic contrast-enhanced (DCE) MRI kinetics can help differentiate ductal carcinoma in situ (DCIS) grades and identify aggressive lesions. Materials and Methods In this retrospective study of patients with DCIS who underwent ultrafast DCE MRI (3.0-9.0 seconds per image) between May 2015 and September 2024, KMC was performed to classify the breast parenchyma, including lesion voxels, into five clusters regarding maximum slope (A × α, where A is initial contrast agent uptake upper limit and α the contrast agent uptake rate). The model-derived and physiologic parameters were computed for each cluster's mean curve and compared with weighted averages from clusters 3-5 across DCIS grades. The Fisher exact, Kruskal-Wallis, Cuzick, and Wilcoxon rank sum tests were performed to identify optimal classifiers for identifying low-grade DCIS and receiver operating characteristic curve analysis to evaluate parameter performance in differentiating between low-grade and intermediate- to high-grade DCIS. Results Among 57 female patients (median age, 52 years; IQR, 20 years; 59 affected breasts), α was associated with different DCIS grades (median α: 2.95 [IQR, 3.01] for low; 5.07 [IQR, 3.63] for intermediate; 6.37 [IQR, 6.33] for high; P = .04, Kruskal-Wallis). The MRI parameters α, A × α, area under the enhancement curve for the first 30 seconds, influx transcapillary transfer constant (K^trans), and efflux transcapillary rate constant (K_ep) had areas under the receiver operating characteristic curve greater than 0.70, with α achieving the highest area under the receiver operating characteristic curve (0.77; 95% CI: 0.56, 0.92; P = .02) for identifying low-grade DCIS. Conclusion KMC-derived kinetic parameters from ultrafast DCE MRI could differentiate low-grade from intermediate- to high-grade DCIS, with α showing the best performance. Keywords: k-Means Clustering, Ultrafast Dynamic Contrast-enhanced MRI, Ductal Carcinoma in situ Grading © RSNA, 2026.