PDA-UNet: A Multiscale Attention UNet for Pathological Image Segmentation

In recent years, pathology has increasingly depended on computational methods to analyze complex biological images. Accurate segmentation of histopathological images is vital for diagnosing and treating various diseases. Modern practice involves scanning H&E-stained tissue sections into high-resolution digital images using Whole Slide Imaging (WSI) technology. In our research, we found that integrating multi-scale correlations within tissue segmentation tasks significantly enhances classification performance when applied to Whole Slide Imaging (WSI) data. Additionally, the incorporation of cross-channel semantic information was found to substantially improve model convergence efficiency. Motivated by these findings, we developed the PDA-UNet architecture, a novel framework that capitalizes on multi-scale feature maps and integrates Squeeze-and-Excitation (SE) Blocks. This design enables the model to more precisely delineate the spatial relationships and boundaries between different tissue regions. Experiments conducted on the BCSS dataset demonstrate that our model achieves significant improvement.