In recent years, two major concerns of mammogram classification tasks are leveraging multi-view mammographic information and class-imbalance handling. In the first issue, multiple multi-view approaches were proposed for combining features from two or more views during training and inference. Overall, most multi-view existing methods are not explainable in the meaning of feature fusion and treat many views equally for diagnosing. Our work aims to propose a simple but novel framework for enhancing the examined view (main view) by leveraging the coarse information from the auxiliary view (ipsilateral view) before learning the comprehensive cancerous features. To address the second issue, we also propose an efficient mammography synthesis framework for upsampling malignant samples. Furthermore, our synthesis framework also has eliminated the limitation of the CutMix algorithm which is unreliable synthesized images with random pasted patches, hard-contour problems, and domain shift problems. Finally, we comprehensively conduct experiments on VinDr-Mammo and CMMD datasets. Our results show that two proposed frameworks for multi-view learning and synthesizing mammography images outperform previous conventional methods in our experiments. • CFDV Net fuses CC and MLO views for accurate cancer classification. • A robust synthesis framework replaces benign regions with malignant cues. • Patch smoothing and Fourier adaptation improve mammogram synthesis quality. • Experiments show strong robustness across fusion settings and datasets.
Nguyen et al. (Wed,) studied this question.