Dynamic tracking of organoids based on RT-DETRv2 and an improved DeepEIoU feature embedding

Organoids have emerged as essential models for precision medicine and drug screening, owing to their high fidelity in replicating the morphology and function of real tissues. However, organoid culture images often have characteristics such as dense distribution, blurred boundaries, and variable shapes, which pose great challenges to automated dynamic tracking. This article proposes a dynamic organoid tracking method that combines the RT-DETRv2 detector with a deep feature embedding module. Firstly, RT-DETRv2 is used for high-precision instance segmentation of organoids in each frame of the image, significantly improving the detection robustness in complex backgrounds. Subsequently, based on the ByteTrack multi-target tracking frame-work, a mask level feature extraction network was introduced, and a DeepEIoU matching strategy that integrates appearance embedding and spatial information was constructed to achieve precise cross-frame identity association of organoids. The proposed approach was evaluated on the custom COCO-Organoid dataset, demonstrating superior performance compared to main-stream methods in multiple indicators such as IDF1 and MOTA, effectively reducing ID drift and mismatch problems, and has good tracking stability and real-time performance. This study provides a reliable technical pathway for high-throughput organoid analysis and long-term tracking.