Heatmap-Seeded Spatially Grounded Query Transformers Improve Astrocyte Soma Detection across Immunostains and Scan Resolutions

Background: Automated astrocyte quantification in human immunohistochemistry (IHC) is hindered by stain variability and dense clustering. Standard Transformers often miss faint somata due to agnostic initialization, while CNNs struggle with overlapping processes. Methods: We propose a Heatmap-Seeded Query Transformer coupling candidate sufficiency with global competitive assignment. Queries initialized from CNN-derived spatial peaks capture faint targets and are refined via a cross-scale Transformer decoder. We evaluated the model on ALDH1L1- and GFAP-stained human tissue (0.35–0.50 μm/px) against Faster R-CNN, YOLOv11, and DETR using COCO metrics and bootstrap-resampled FROC analysis. Results: Our method consistently improved AP across all cohorts, notably by ∼20% in highresolution ALDH1L1 images relative to baselines. FROC analysis confirmed superior sensitivity at low false-positive rates, validating effective noise suppression with high recall. Conclusions: Integrating spatial priors with global reasoning mitigates the sensitivity-precision trade-off, offering a robust tool for clinical neuropathology resilient to domain shifts in archival tissue.