Multi-pathway fusion network for early-stage breast tumor segmentation from MRI

Breast cancer is one of the most common cancers in women, with a notably high mortality rate. Early diagnosis can improve survival rates. However, early-stage breast tumors suffer challenges for accurate detection and are hard to detect due to their tiny sizes and blurry edges, thereby obtaining degraded performance.

To solve the above issues, this study aims to develop a robust model for the early-stage breast tumor segmentation from magnetic resonance imaging (MRI) and to provide a quality assessment for early-stage breast cancer.

We propose an early-stage breast tumor segmentation method named MPNet, which utilizes a multi-pathway fusion strategy, focusing on preserving tumor boundary information while processing their contextual information. Our approach consists of two main pathways: the detail information pathway (DIP) and the context enhancement pathway (CEP). The DIP preserves the tumor boundary details by capturing high-resolution features, while the CEP enhances the semantic information by enlarging the receptive field and introducing quarter-scale global self-attention for global contextual information. We also design a bilateral feature fusion module to fuse the representations from different pathways, facilitating interaction between both types of features. Additionally, we collect a clinical dataset for early-stage breast cancer diagnosis, comprising 260 diverse cases.

Comparative experiments show the effectiveness of our method on clinical data, where the mean intersection over union and Dice similarity coefficient are 87.41% and 85.69%, respectively.

Overall, MPNet demonstrates satisfying performance on segmenting early-stage breast tumors with tiny sizes by preserving boundary details and enhancing contextual information. Extensive experiments demonstrate that MPNet outperforms state-of-the-art methods for enhancing early breast cancer intervention.