IHA-YOLO: Inter-Head Attention for Real-Time Cell Detection

Multiclass cell detection is a crucial task in numerous biomedical applications, particularly in cell biology. The development of YOLO object detection models has advanced the field of real-time detection, but it is still struggling with challenges in medical imaging due to limited data availability, overlapping tiny objects, diverse cell types, and class imbalances. In this paper, we introduce Inter-Head Attention (IHA)-YOLO, a novel model that proposes an inter-head self-attention block to enhance global representation learning, thereby improving the contextual understanding across feature maps and performing effective detection of small cells and sub-cell structures in medical images. Through extensive experiments on five publicly available datasets, IHA-YOLO outperforms the state-of-the-art methods, achieving an average absolute mAP50 improvement of 2.03% and a 13% faster inference rate. In addition to cell detection, we adapt IHA-YOLO for cell counting to demonstrate its effectiveness. Our source code and models are available at: https://github.com/toobatehreem/IHA-YOLO.