A deep learning model integrating multi-phenotypic features of microcalcifications enhances malignancy prediction for BI-RADS 4 microcalcifications in mammography: a multicenter study

Abstract

Some Breast Imaging Reporting and Data System (BI-RADS) 4 category microcalcifications (MCs) exhibit atypical or overlapping characteristics in terms of morphology and distribution, posing a diagnostic challenge for doctors. The aim was to develop and evaluate a deep learning (DL) model for predicting the malignancy of BI-RADS 4 MCs based on mammography. This retrospective study collected 1708 patients from two centers based on inclusion/exclusion criteria, dividing the cohort into training, validation, internal and external testing cohorts. After image segmentation, the phenotypic features of MCs, including semantic, morphological radiomics, deep convolution, and topological features, were extracted, and a BMC MG -Net hybrid framework, merging a convolutional neural network (CNN) and graph convolutional network (GCN) for predicting MCs type was developed. The area under the receiver operating characteristic curve (AUC) analysis was utilized to quantitatively assess and compare diagnostic efficiency among different model architectures, junior and senior radiologists, and radiologists employing the model. The BMC MG -Net model achieved AUCs of 0.86 and 0.87 in the internal test ( n = 308) and external test cohort ( n = 163), respectively. The proportion of lesions downgraded by this model was 20% (61 of 308) and 40% (65 of 163) for BI-RADS 4 in these two cohorts, respectively. With the assistance of this model, the AUC of the junior and senior radiologists increased to 0.80 and 0.89 from 0.68 to 0.82, respectively. BMC MG -Net based on mammography can predict the malignancy of MCs, which suggests that this method can help reduce unnecessary biopsies for benign patients.