BCNet: integrating UNet and transformer for blood cell segmentation

Automatic segmentation of blood cells is crucial in medical diagnosis and research, significantly improving the accuracy and efficiency of diagnosing blood disorders. Traditional segmentation methods involving manual segmentation are time-consuming, labor-intensive, and prone to errors. In recent years, advancements in deep learning have provided new solutions for automated segmentation. This paper proposes BCNet, a blood cell segmentation algorithm combining UNet and Transformer. Specifically, BCNet utilizes UNet’s Encoder-Decoder architecture as the backbone for extracting multi-scale features. A Spatial Reduction Transformer (SRT) Module is introduced for capturing long-range dependencies in the deepest downsampling layers to enhance sensitivity to local features. Additionally, coordinate attention is employed instead of skip connections for multi-scale feature fusion, enriching semantic information in deep features. Experimental results demonstrate that BCNet achieves superior Dice and IoU metrics compared to classical medical image segmentation models, facilitating automated analysis and medical diagnosis of blood cells.