A Multi-Strategy Active Learning Framework for Enhanced Peripheral Blood Cell Image Detection

Peripheral blood cell detection is essential for diagnosing and monitoring hematologic disorders. However, existing methods typically focus on a limited number of cell types (usually 3 to 10), restricting their ability to detect a broader range of cells. These methods also struggle with class imbalance and still rely heavily on large amounts of annotated data, limiting their effectiveness for underrepresented categories and hindering scalability. To address these challenges, we propose a novel composite multistrategy active learning framework using YOLOv5 for enhanced peripheral blood cell detection. The framework reduces annotation costs and improves detection performance by combining uncertainty-based selection, diversity querying, and density-based querying to prioritize the most informative and diverse samples. The process begins with entropy-based uncertainty selection to identify the most uncertain samples, followed by clustering analysis to capture diverse samples from the feature space, and concludes with densitybased selection using the k-nearest neighbors algorithm to prioritize samples from high-density regions. Experimental results demonstrate that the framework achieves a mean average precision ([email protected]) of 64.8% on a private dataset with 26 cell types, outperforming other active learning strategies and existing methods. It also reduces manual annotation workload by 28.7% compared to random sampling. On the public BCCD dataset (3 cell types), the framework achieves an [email protected] of 86.7%. These results highlight the practicality and reliability of the proposed framework for optimizing peripheral blood cell detection.