Update of Breast MR Imaging Architectural Interpretation Model

Publisher Website
Google Scholar
Abstract
To (a) validate a breast magnetic resonance (MR) interpretation model, (b) expand the tree-shaped prediction model to increase specificity without decreasing sensitivity, and (c) reevaluate the model's diagnostic performance. Two hundred sixty-two new patients with palpable or mammographic abnormalities underwent MR imaging, and pathologic evaluation was performed. They were entered prospectively into the model, which yielded 454 patients in the construction (training) and validation (test) phases. Predictive values for previously published terminal nodes or branch points of the model were compared between the training and test data sets. Ductal enhancement morphology, regional enhancement micronodularity, regional enhancement degree, and focal mass T2 signal intensity were evaluated for model expansion. Diagnostic performance characteristics of the model were recalculated. For previously published nodes, absence of a lesion visible at MR imaging, smooth masses, lobulated masses with nonenhancing internal septations, and lobulated masses with minimal or no enhancement had negative predictive values (NPVs) for malignancy similar in both data sets (96% vs 99%, 100% vs 93%, 100% vs 98%, and 100% vs 100%). Irregular masses with internal septations (100% vs 0%) and spiculated masses with no or minimal enhancement (100% vs 50%) did not. Nonseptated enhancing lobulated masses with low T2 signal intensity were added as a benign terminal node (NPV, 100%). Mild regional enhancement (NPV, 92%) was added but not considered a terminal node. Sensitivity, specificity, NPV, positive predictive value, and accuracy of the expanded model were 96%, 80%, 96%, 78%, and 87%, respectively. Additional investigation yielded a slightly modified model, but the diagnostic performance characteristics remain high, similar to those originally published.