FSA-UNet: A four-stage-attention UNet based semantic segmentation for breast cancer detection

Abstract

Author(s): Pothina Praveena* and N. Suresh Kumars

Breast cancer is a severe global health concern, accounting for a significant portion of the annual mortality rate among women. It now ranks first among female cancer cases globally, having a significant impact on both healthcare systems and society at large. This demonstrates clearly that early detection, precise diagnosis and appropriate disease management can all significantly improve the functioning of the healthcare facility. To address this, we offer a multi-attention-based UNet (FSAUNet). In particular, we suggest a residual encoder that enhances the backbone's fine-grained feature extraction capacity. It is built on a basic attention module. The underlying semantic representation of the provided feature map is reconstructed by applying multi-head self-attention on the lowest level feature, hence enabling fine-grained segmentation for various pixel categories. Then, to more accurately merge the feature data of the target across various scales, we employ channel attention and spatial attention during various feature fusion stages. This helps to address the issue of multiple scales in various categories. We also increase the number of down samplings to divide the target's feature sizes at different scales.

Share this article