MammoDenseSegNet outperforms VGG16 baseline for dense tissue segmentation across mammography devices

New encoder-decoder CNN (MammoDenseSegNet) segments breast dense tissue more reliably than VGG16 baseline across 3 datasets (1,499 mammograms, 606 women): Dice 0.63–0.91 vs 0.06–0.82 (p<0.001). Biggest gain in low-density breasts where baseline nearly fails (Dice 0.16→0.63).

• Retrospective diagnostic-accuracy study; 1,499 mammograms from 606 women across two public datasets (VinDR-Mammo, EMBED) and one private dataset spanning varied densities and imaging artifacts.
• Architecture combines adaptive dual-attention module (spatial + channel), multi-kernel receptive field module, and multi-scale Dice loss with deep supervision; tested against a publicly available VGG16-based segmentation algorithm.
• Key limitation: relatively small, single-sex cohort (606 women); no prospective or external multi-institutional validation beyond the three included datasets, limiting generalizability to other scanner vendors and clinical workflows.

Automated summary

RadPigeon summaries are original and for information only. They are not clinical advice.