Blood Cell Detection and Self-Attention-Based Mixed Attention Mechanism

Deep learning-based microscopy image analysis has made significant progress. However, accurately identifying cell targets among dense and complex distributions is still challenging. This study introduces a single-stage anchor-free Blood Cell Detector (BCDet) to address the low recognition rate caused by this circumstance, including a novel self-attention-based mixed attention mechanism that sequentially integrates modeling information from channel dimension and spatial dimension. The first part selectively emphasizes interdependent channel maps by integrating associated features among all channel maps. The second part selectively aggregates the feature at each position by a weighted sum of the features at all positions. Additionally, the MIFI(Multi-Head Attention-based Intra-scale Feature Interaction) block is introduced to the processing of upper and lower feature data to improve the network's capacity to gather details about various cell properties, enabling BCDet to better recognize cells in cell-dense areas. Experimental results on the BCCD dataset demonstrate that our method achieves mAP@0.5:0.95 of 67.0%.