Semantic Segmentation of Mammography Images with Deep Learning Derin grenme ile Mamografi G r nt lerinin Anlamsal B l tlenmesi

Emiroglu M. B., Kalamanoglu A., BİLGİN G.

2025 Innovations in Intelligent Systems and Applications Conference, ASYU 2025, Bursa, Türkiye, 10 - 12 Eylül 2025, (Tam Metin Bildiri) identifier

  • Yayın Türü: Bildiri / Tam Metin Bildiri
  • Doi Numarası: 10.1109/asyu67174.2025.11208474
  • Basıldığı Şehir: Bursa
  • Basıldığı Ülke: Türkiye
  • Anahtar Kelimeler: Breast cancer, data augmentation, deep learning, mammography segmentation, medical image analysis
  • Yıldız Teknik Üniversitesi Adresli: Evet

Özet

In this study, different architectures for deep learning-based segmentation of mammography images were compared. The DMID dataset, containing 510 high-resolution images, was augmented using horizontal flipping, rotation, and scaling due to its limited data size and imbalanced distribution, and then divided into training, validation, and test sets. Only abnormal images were used to enhance the learning of small lesions. U-Net, Attention U-Net, ResUNet, MultiResUNet, DeepLabV3, and EfficientViT models were evaluated with hyperparameter and loss function optimization. The combination of Weighted Binary Cross-Entropy (WBCE) and Dice loss provided the best results against class imbalance. Evaluated using Dice, Precision, Sensitivity, IoU, and Cohen's Kappa metrics, the MultiResUNet model achieved the highest performance with a Dice score of 79.13%. The results indicate that architectural selection and loss function optimization are crucial for segmentation success.