Purpose: To identify a model that effectively classifies suspicious breast lesions based on kinetic features measured on high-temporal resolution (ultrafast) DCE-MRI.
Methods: 140 women were imaged on a 3T scanner with a protocol that included ultrafast DCE-MRI with a temporal resolution of 3 – 9 seconds per frame for roughly one minute following injection of contrast media. A total of 204 lesions (129 malignant and 75 benign) were included in the analysis. Relative signal enhancement vs time curves were fit to an empirical mathematical model. From the fits the following features were calculated: time to initial enhancement (TTE), time to 20% relative enhancement (t20), maximum slope of enhancement (MS), time to maximum slope of enhancement (tMS), relative enhancement at 30 seconds following bolus arrival in the aorta (SE30), and initial area under the uptake curve (iAUC). Features were measured as averages for the whole enhancing lesion ROI and the average value in the top 15% enhancing voxels (hotspot analysis). Times for each case were expressed relative to the bolus arrival time in the aorta, reducing inter-patient variability, e.g. due to cardiac output. The binary outcome for each lesion (benign or malignant) along with all the features measured were used to fit a logistic regression model. Leave-one-out cross validated LASSO regression was then used to perform feature selection and identify the most parsimonious model for lesion classification.
Results: The model identified by lasso regression included 3 kinetic parameters: SE30, tMS, and t20 – all measured in the top 15% of enhancing voxels in the lesion. This model achieved an area under the ROC curve of AUC = 0.82 [95% CI: 0.76-0.87].
Conclusion: Initial enhancement measured on ultrafast DCE-MRI is an effective classifier for suspicious breast lesions. Lesions that enhance rapidly and to a greater extent are more likely to be malignant.