Abstract
OBJECTIVE: To evaluate the positive predictive value (PPV) of the imaging characteristics of breast lesions classified as BI-RADS category 4 (risk of malignancy > 2% to < 95%) on magnetic resonance imaging (MRI), in order to create an algorithm to subcategorize such lesions. MATERIALS AND METHODS: This was a retrospective study including 199 breast lesions (131 nodules and 68 non-mass lesions) classified as BI-RADS 4 on MRI. Of the 199 lesions, 93 were excluded, for various reasons: they were lymph nodes; they were not biopsied or were not followed at our center; they were additional findings in patients with an established diagnosis of malignancy who underwent mastectomy without further investigation; or they were identified in examinations that were not recovered from the digital archive. Multivariate analysis was performed to identify the most relevant descriptors to predict malignancy and to build an algorithm to subcategorize lesions into BI-RADS 4A, 4B, and 4C. Four breast radiologists then tested the algorithm in another 95 patients with breast lesions classified as BI-RADS 4 on MRI, 27 (28.4%) of those lesions having previously been classified as malignant. RESULTS: The descriptors statistically associated with malignancy in the multivariate analysis of the nodules were background parenchymal enhancement, margins, and the initial phase of the kinetic curve. An algorithm was developed by using these resources, and the PPV obtained for each category was 4.3% for BI-RADS 4A, 21.4% for BI-RADS 4B, and 78.9% for BI-RADS 4C. In the validation of the algorithm by the four breast radiologists, the PPV of the subcategories was within the BI-RADS malignancy ranges in almost all situations, the exceptions being the 11.1% that one evaluator obtained for category 4A and the 46.4% obtained for category 4C by another evaluator. CONCLUSION: The objective analysis employing the proposed algorithm proved useful for subdividing BI-RADS 4 mass lesions on MRI and showed better interobserver agreement than did the subjective analysis.