GGC-DETR: Morphology-Sensitive Lightweight Blood Cell Localization for Automated Peripheral Blood Smear Review

Peripheral blood smear review remains important when automated hematology analyzer outputs require morphological confirmation. Automated cell localization can extract red blood cell, white blood cell, and platelet regions for downstream counting, morphology assessment, and review support. However, dense erythrocyte backgrounds, small platelet targets, inter-class scale variation, and constrained deployment settings make accurate lightweight localization difficult. This study proposes GGC-DETR, a lightweight RT-DETR-R18-based detector organized around morphology-sensitive localization under constrained computation. Morphology sensitivity is treated as preservation of contour, boundary, platelet-scale, and nuclear-cytoplasmic evidence, not as additional semantic labeling. The architecture combines global width scaling, Ghost-based backbone reconstruction, CabC3 local representation, and selective GSConv neck fusion to reduce redundant computation while retaining local and multi-scale evidence. The model was evaluated on two public peripheral blood smear datasets under a unified protocol using detection accuracy and efficiency metrics. On BCCD, GGC-DETR achieved an mAP $_{50}^{}$ of 0.921, an F1-score of 0.884, and a Recall of 0.946 with 5.2 M parameters, 16.9 GFLOPs, a 24.1 MB model size, and 91.4 FPS. On TXL-PBC, the same configuration achieved an mAP $_{50}^{}$ of 0.979, an F1-score of 0.974, and 92.7 FPS. Relative to RT-DETR, GGC-DETR increased BCCD mAP $_{50}^{}$ by 2.4 percentage points while reducing parameters and GFLOPs by approximately 74% and 70%, respectively, defining a compact high-recall operating point for automated peripheral blood smear image analysis.