Automated cell detection and identification of mitotic division events in phase-contrast microscopy are essential for quantifying proliferation, phenotypic response, and population dynamics, yet practical deployment remains limited by annotation cost and unclear architecture selection under real-time constraints. This paper releases B16BL6, a public phase-contrast time-lapse microscopy dataset comprising 1800 frames (1600 × 1200) with 93,637 bounding-box annotations for two classes (Cell and Division). To enable scalable labeling at this density, we present a three-stage annotation pipeline that combines (i) expert manual labeling of a representative subset, (ii) YOLO-based pseudo-label generation with expert refinement, and (iii) temporally consistent label propagation using a detector–tracker hybrid with an enhanced DeepSORT variant. Using this dataset, we benchmark seven contemporary detectors spanning one-stage CNNs (YOLOv8x, YOLOv11-L, SSD300/512), two-stage models (Faster R-CNN, Mask R-CNN, Libra R-CNN), and a real-time transformer (RT-DETR-L). Evaluation is multi-dimensional, covering detection accuracy (mAP@0.50 and mAP@0.50:0.95), class-specific AP, inference speed, parameter/GFLOP complexity, and imbalance sensitivity via the Cell-Division AP gap, together with a supplementary heuristic efficiency score for deployment-oriented comparison. Results show that YOLOv8x achieves the best accuracy (mAP@0.50 = 0.837), YOLOv11-L provides the best accuracy–speed balance (20.15 FPS, mAP@0.50 = 0.811), and two-stage detectors exhibit more stable class-level performance under imbalance, while SSD offers maximal throughput at reduced precision. Rather than proposing a new detection architecture, this work releases a densely annotated microscopy dataset, a scalable annotation workflow, and a unified benchmark of modern detectors, providing practical guidance for scalable microscopy pipelines and downstream tracking and division analysis.
Al-Hamadani et al. (2026) studied this question.