Lesion-Aware Image Augmentation for Robust Mammogram Classification

Breast cancer is one of the leading causes of cancer-related mortality among women worldwide, making early and accurate mammogram classification essential for its effective diagnosis and treatment. Data augmentation has been widely employed to improve the generalization capability of deep learning models; however, conventional augmentation methods may distort diagnostically important lesion structures or suppress critical high-intensity regions. In this study, a Lesion-Intensity Aware Augmentation (LIAA) method is proposed for mammogram classification. The proposed approach extracts a local patch from the original mammogram and pastes it into a randomly selected target position while preserving the original image’s structure outside the modified region. During the blending process, intensity-guided fusion is performed by selecting the maximum intensity value between the source patch and the corresponding target region at each pixel location. This strategy introduces localized variability while preserving high-response lesion structures and reducing the risk of lesion suppression after patch replacement. The effectiveness of the proposed method was evaluated using the CBIS-DDSM and MIAS datasets with deep learning models for benign–malignant classification. The experimental results demonstrate that LIAA achieved superior classification performance compared with conventional augmentation techniques. In particular, the proposed method obtained the highest AUC score and improved the discriminative capability of the classification models by generating more informative lesion representations while maintaining diagnostically relevant structures.