YOLO-AF: An Improved Lightweight Architecture for Real-Time Cell Detection in Imaging Flow Cytometry

The imaging flow cytometer enables real-time capture of microscopic cell images under dynamic conditions of continuous high-speed flow, thereby acquiring instantaneous morphological and structural information of cells in motion. This significantly enhances the information dimensionality and feature integrity of cellular feature data. Detecting and identifying different types of flow cytometry images facilitates early disease screening and auxiliary diagnosis. To enhance the detection efficiency and accuracy of cells in flow imaging, this paper proposes an improved model named YOLO-AF for real-time detection of high-speed flowing cell images. Based on YOLOv12, our model integrates two key improvements: First, the lightweight attention module AFE (Attention Feature Enhancement) is introduced. Combining channel and spatial attention mechanisms, AFE adopts adaptive compression and spatial information aggregation strategies to enhance the model's ability to extract features of small cells. Second, the Feature Pyramid Network (FPN) is reconstructed to further improve multi-scale feature fusion, thereby enhancing detection performance for cell targets of varying sizes. Additionally, this study constructs a high-resolution microscopic image dataset of animal cells, named AnimalRBC, using a lab-made imaging flow cytometer. The dataset includes three types of red blood cells: guinea pig, chicken, and sheep. Experimental results show that YOLO-AF achieves 95.5% mAP@0.5 on this dataset, a 3.2% improvement over the baseline model YOLOv12. This method significantly enhances recognition accuracy while maintaining high detection speed, providing robust support for cell-related early disease screening.