RexGANet based classification method for microcalcifications in mammograms

Breast microcalcifications are one of the important indicators of early breast cancer. Radiologists discriminate benign and malignant microcalcifications mainly based on their morphology and distribution. Convolutional neural network (CNN) excels at extracting morphological features of objects, while cannot accurately describe the spatial structural relationships among microcalcifications. To address this issue, a novel microcalcification classification model (RexGANet) that integrates CNN and graph attention network is proposed. In this model, the improved ResNeXt101 network is used to extract morphological features of microcalcifications, while the multihead graph attention network is utilized to extract their spatial distribution features. These two types of features are extracted by a dual branch parallel network and fused through the improved attentional feature fusion module to achieve microcalcification classification based on morphological and distribution features. In addition, to address the issue of class imbalance between benign and malignant microcalcifications, large margin-aware focal loss (LMFLOSS) is used to replace the cross-entropy loss function. The experimental results demonstrate that RexGANet achieves classification accuracies of 94.46% and 92.53% for DDSM and BCDR datasets, respectively, outperforming the state-of-the-art methods in the same field.