Edge-Texture and Transformer Enhanced Dual-Path Networks for Breast Mass Segmentation in Mammograms
6th International Conference on Medical Imaging and Computer-Aided Diagnosis, MICAD 2025, London, İngiltere, 19 - 21 Kasım 2025, cilt.1520 LNEE, ss.127-137, (Tam Metin Bildiri)
- Yayın Türü: Bildiri / Tam Metin Bildiri
- Cilt numarası: 1520 LNEE
- Doi Numarası: 10.1007/978-981-95-7425-4_12
- Basıldığı Şehir: London
- Basıldığı Ülke: İngiltere
- Sayfa Sayıları: ss.127-137
- Anahtar Kelimeler: Breast mass segmentation, Cross-attention, Deep learning, Mammography, Transformers, U2-Net
- Ankara Yıldırım Beyazıt Üniversitesi Adresli: Evet
Özet
Accurate segmentation of breast masses in mammograms is essential for early detection and effective treatment of breast cancer. We propose three lightweight dual-path architectures—ETDP-U2-Net, DPCA-U2-Net, and DPTrans-U2-Net—that integrate edge–texture fusion, cross-attention, and transformer-based encoding to enhance feature representation while drastically reducing model size. On the CBIS-DDSM dataset, our models consistently outperform U-Net variants despite being up to three times smaller. ETDP-U2-Net achieved the best performance (Dice: 0.9162, IoU: 0.8485), while DPTrans-U2-Net delivered comparable accuracy with only 4.2M parameters. These results highlight the potential of the proposed architectures to combine efficiency and accuracy, making them promising candidates for real-world computer-aided diagnosis in breast cancer screening.