BMA-YOLO: an object detection model for microscopic images of mouse embryos

In vitro embryo culture is a pivotal technology in life sciences and medical research. However, automated monitoring remains challenging due to factors such as bubble interference and the frequent omission of small or peripheral embryos. To overcome these limitations, this study introduces BMA-YOLO, an enhanced model built upon the YOLOv8 architecture. By integrating BiFormer and Multi-Scale Attention Aggregation (MSAA) mechanisms, the model strengthens its capacity for feature extraction and multi-level feature fusion. Furthermore, a detection head specifically designed to capture small objects is incorporated to improve sensitivity toward diminutive embryos. Experimental results on a private embryo dataset demonstrate that BMA-YOLO achieves an mAP50 of 96.94%, an mAP50-95 of 69.26%, and a recall rate of 94.20%. Compared with the baseline model, these results correspond to improvements of 3.57%, 3.30%, and 6.69%, respectively. BMA-YOLO effectively mitigates bubble interference and significantly enhances the detection performance for small and peripheral embryos, providing robust algorithmic support for subsequent automated monitoring systems.