Topology-Guided ACO–YOLOv8n: A Metaheuristic-Driven Hyperparameter Optimization Framework for Blood Cell Detection and Classification

Automated blood-cell detection can support microscopic analysis, but detector performance depends strongly on the training configuration. This study investigates a topology-guided Ant Colony Optimization (ACO) strategy for jointly selecting 21 hyperparameters of YOLOv8n without modifying its architecture. This study uses 614 images from BCCD and LISC, with 5138 available annotations for red blood cells (RBCs), white blood cells (WBCs), and platelets. Random Search, Standard ACO, and topology-guided ACO were compared within the same discretized search space and with a maximum of 600 unique candidate evaluations per run. Across five reported runs, topology-guided ACO reached a mean best validation mAP@0.5 of 96.8±1.2%, compared with 94.3±1.6% for Standard ACO and 91.6±2.2% for Random Search. The selected detector achieved 94.4% precision, 93.3% recall, 95.8% mAP@0.5, and 62.3% mAP@0.5:0.95 on the reported held-out evaluation. Because LISC provides WBC-specific annotations rather than complete three-class labels, source-specific interpretation is required. The present contribution is therefore methodological: this study assesses whether transition-based search information can improve systematic hyperparameter selection for lightweight microscopic object detection.