Multi-view deep learning approach improves breast cancer detection rates in mammography

Researchers have developed a systematic framework for optimizing breast cancer detection in mammographic imaging using modern machine learning techniques. The work addresses critical gaps in how AI systems analyze chest X-rays by examining five core design questions in diagnostic model development.

Methodological approach#

The study investigated how transfer learning from natural images applies to medical imaging, compared conventional downscaling approaches with adaptive resizing techniques, and examined the benefits of simultaneously processing multiple mammogram projections from different angles. The team benchmarked these approaches against two widely-used datasets—CBIS-DDSM and the complete VinDr-Mammo collection—while also testing whether performance remains consistent across varying image quality levels.

Key findings#

On the CBIS-DDSM dataset, the researchers achieved notable improvements: single-view analysis improved from 0.8153 to 0.8343 AUC, while multi-view detection rose from 0.8483 to 0.8658 AUC. Using a novel comparative method, extending the system from single to multiple viewpoints yielded a 0.0217 AUC gain.

Performance gains were even more pronounced on the complete VinDr-Mammo dataset, where the multi-view approach delivered a 0.0492 AUC increase relative to single-view analysis, reaching an overall 0.8511 AUC score.

Implications for clinical screening#

These advances establish new reference standards for mammogram analysis and provide clear evidence that jointly processing complementary imaging angles produces more reliable diagnostic systems. The findings also contribute principles for transfer learning and architectural design that could guide development of more dependable breast cancer screening tools.

The team has publicly released inference code and trained models to support further research in this domain.