Akademik Veri Yönetim Sistemi
Biomedical Signal Processing and Control, cilt.110, 2025 (SCI-Expanded)
Breast cancer is the most common cause of death among women. Accurate and early detection from histopathological images utilizing computerized methods greatly enhances the chances of survival. However, developing an automated method for this purpose is challenging due to many factors, including complex variations between tumor classes, irrelevant features, inconsistencies in stain colors, and limited datasets. In this work, a new approach that overcomes these challenges has been developed. The main contributions of this work are introducing a stain-identification-based, unsupervised stain normalization technique and a hybrid network specifically developed for breast cancer detection. The network has been intelligently designed by utilizing the strengths of vision transformers and convolutional networks. It contains two new adjustable modules, effectively extracting both global contextual and detailed information. One of them is an adjustable convolution module, which comprises dilated and separable convolution layers and dynamically extracts rich features at multiple scales. The other module is an adjustable attention module, which captures complex patterns and long-term dependencies and enhances learned representations by utilizing a dynamic multi-head attention mechanism. Additionally, comprehensive experiments and fair comparisons of recent methods have been conducted using the same datasets and identical metrics. The images have been provided from the public breast cancer histopathology image database. Experimental results demonstrate that the proposed method surpasses recent methods in binary classification of breast cancer, with 98.76% accuracy, 98.56% recall, 97.63% precision, and a 98.49% F1-score. It also performs better in multi-class classification, with 96.58% accuracy, 96.69% recall, 96.51% precision, and a 96.64% F1-score.