BCRT-DETR: A Lightweight Blood Cell Detection Transformer Model for Enhanced Medical Diagnostics.

Abstract

In the biomedical field, the detection of microscopic images of blood cells is crucial for diagnosing of blood-related diseases. To enhance accuracy and real-time performance, we developed a blood cell real-time detection transformer (BCRT-DETR) to improve detection efficiency. A dynamic alignment integration backbone network (DAIBN) was introduced to address the spatial differences in features from diverse sources during multi-backbone information fusion. Additionally, a multi-scale parallel aggregation splicing (MPAS) module was integrated into the neck component to mitigate missed detections during cell feature extraction. The integration of high- and low-frequency (HiLo) attention with the attention-based intra-scale feature interaction (AIFI) module to form AIFI-HiLo effectively overcame the model's previous limitation of concentrating on regions with cellular density. The introduction of the retentive meet transformer block (RMT_Block) in the neck component further optimized the computational complexity, thereby increasing the detection speed. The experimental results indicated that, compared with the recent transformer-based real-time detection model, RT-DETR, BCRT-DETR achieved significant efficiency improvements with reductions of 33.8%, 51.1%, and 34.1% in parameters, giga floating-point operations per second (GFLOPs), and model size, respectively. Simultaneously, BCRT-DETR improved mAP50 and mAP50:95 by 0.8% and 1.6%, respectively, with mAP50 reaching 96.8%. Furthermore, BCRT-DETR demonstrated exceptional generalization capabilities across the blood cell detection, blood cell count and detection, and complete blood count datasets. Our model provides reliable technical support and offers innovative solutions for automated medical diagnosis.