Leveraging Dual-Teacher Collaboration and Discriminative Language Guidance for Robust Blood Cell Detection

Automated blood cell detection (BCD) is essential for modern hematology, yet its robustness is challenged by the limitations of fully supervised models. The impracticality of exhaustively annotating the vast number of cells in blood smears leads to extremely sparse annotations and, consequently, high rates of missed detections. Moreover, the clinical utility of current models is severely impeded by poor generalization, a result of domain shifts across clinical sources and a restrictive closed-set assumption. To overcome these issues, we first propose a dual-teacher collaboration framework that employs a hierarchical fusion-filtering strategy to integrate the strengths of both the local teacher and the external teacher, effectively counteracting missed detections from sparse data. Furthermore, we leverage language guidance to encourage the detector to learn visual representations with stronger cross-domain generalizability, and introduce a Multi-Contrastive Discriminability Calibration (MCDC) framework that achieves controllable inter-class separation, mitigating the decline in discriminability introduced by language guidance. We evaluate the pseudo-label generation strategy on an extremely sparsely annotated dataset, UniBCD-24, and achieve average recall improvements of 1.3, 51.9, and 11.1 for nucleated cells, erythrocytes, and platelets, respectively. Additionally, we instantiate the MCDC framework with YOLO-World, yielding MCDC-YOLOW, which achieves state-of-the-art fully supervised detection performance alongside superior zero-shot domain-transfer ability across diverse visual domains, demonstrating a favorable trade-off between strong generalizability and high discriminability.