Breast X-ray Image Four-Class Classification Based on Channel Attention and Tri-Channel Pseudo-Color Input

With the rapid development of deep learning in the field of medical imaging and the increasing incidence of female breast diseases, the classification and diagnosis of mammographic images have become an important research direction in the computer field. To address the problems of insufficient utilization of channel information and low contrast sensitivity in X-ray image feature extraction by traditional convolutional neural networks, this paper proposes a mammographic image classification method combining the Efficient Channel Attention (ECA) module and Tri-Channel Enhancement (Tri-CE). The CBIS-DDSM public dataset was adopted, and a four-classification model was constructed based on DenseNet121, including benign calcification (BEN_CALC), malignant calcification (MAL_CALC), benign mass (BEN_MASS), and malignant mass (MAL_MASS). Systematic experiments were conducted under strategies such as pre-training, WarmCos learning rate, and EMA smoothing. Experimental results show that the base model achieves an accuracy (Acc) of 0.663 and an F1-score of 0.649, while the model integrated with the ECA module and Tri-Channel Enhancement yields an Acc of 0.695 and an F1-score of 0.681, demonstrating slightly better stability and generalization ability. This study provides a lightweight and scalable feature enhancement approach for the automatic diagnosis of breast imaging.