Microcalcification Classification in Mammograms Based on HyperResNeXt

Microcalcifications in the breast are one of the important signs for the early diagnosis of breast cancer. Accurate differentiation between benign and malignant microcalcifications is of great significance for improving the early diagnosis rate of breast cancer. This paper proposes a novel microcalcification classification model named HyperResNeXt, which integrates convolutional neural networks (CNN) and hypergraph convolutional neural networks (HGCN). A dual-branch network structure is used to extract the morphological and spatial distribution features of microcalcifications. The CNN branch is employed to extract deep features of microcalcifications in mammograms, and these features are used to construct their hypergraph representations to enhance the ability to characterize and mine the nonlinear high-order association between samples. The HGCN branch is designed for the hypergraph representation to extract the spatial distribution features of microcalcifications. An improved attention-based fusion module is used to implement feature fusion, thereby achieving the classification of benign and malignant microcalcifications based on the fused features. To address the class imbalance problem between benign and malignant microcalcification samples, the Large Margin aware Focal Loss (LMFLOSS) is used instead of the traditional cross-entropy loss, effectively optimizing the model’s learning ability for minority class samples. The effectiveness of the proposed method is evaluated using the benchmark DDSM database. Experimental results show that HyperResNeXt achieves a classification accuracy of 94.71% and significantly outperforms existing CNN methods. This provides an efficient approach to the automated classification and non-invasive diagnosis of breast microcalcifications.