Heatmap-seeded spatially grounded query Transformers improve astrocyte soma detection across Immunostains and scan resolutions
- 1Early Intervention Unit, Department of Psychiatry, Affiliated Nanjing Brain Hospital, Nanjing Medical University, Nanjing, China
- 2School of Biomedical Engineering and Informatics, Nanjing Medical University, Nanjing, Jiangsu China
Abstract
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 instances.
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 \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}(0.35−−0.50 μm/px)\end{document} against Faster R-CNN, YOLOv11, and Deformable DETR using COCO-style metric under the Olar et al. evaluation protocol, bootstrap-resampled FROC analysis, and targeted stratifications for local contrast and pair-separation recall.
Results: Our method improved overall Average Precision (AP) across the vast majority of test cohorts, including a nearly 19% relative increase in high-resolution ALDH1L1 images. Crucially, stratified evaluations revealed significant improvements in localizing low-contrast targets and resolving densely crowded cell pairs without relying on heuristic filtering. FROC analysis confirmed superior sensitivity at clinically relevant low false-positive rates.
Conclusions: Integrating spatial priors with global reasoning effectively mitigates the sensitivity-precision trade-off. This yields a robust detector for astrocyte soma localization resilient to stain and resolution shifts, serving as a highly reliable upstream component for quantitative digital pathology workflows.
Supplementary Information: The online version contains supplementary material available at https://doi.org/10.1186/s12880-026-02678-5.
Related
- G Protein-Coupled Receptors Across Diverse Astrocyte Functions and Cellular Crosstalk
- Neuro-regulator role of H2S in astrocyte activation and its effects on neurological damage and behavior of VPA-exposed rats
- Endogenous Neuronal-Like Phenotypic Remodeling of Astrocyte-Associated Cells in the Ischemic Cortex After Stroke
- Human astrocyte aging couples chromatin closure to focal G-quadruplex accumulation
- Vision Transformers as a Framework for Imaging Fluorescence Correlation Spectroscopy
BibTeX
@article{zhu2026heatmapseeded,
title = {{Heatmap-seeded spatially grounded query Transformers improve astrocyte soma detection across Immunostains and scan resolutions}},
author = {Jiayang Zhu and Xizhe Zhang},
journal = {BMC Medical Imaging},
year = {2026},
month = aug,
doi = {10.1186/s12880-026-02678-5},
url = {https://doi.org/10.1186/s12880-026-02678-5},
note = {CC BY-NC-ND 4.0},
}
RIS
TY - JOUR
TI - Heatmap-seeded spatially grounded query Transformers improve astrocyte soma detection across Immunostains and scan resolutions
AU - Jiayang Zhu
AU - Xizhe Zhang
PY - 2026
DA - 2026/08/17
JO - BMC Medical Imaging
DO - 10.1186/s12880-026-02678-5
UR - https://doi.org/10.1186/s12880-026-02678-5
AB - 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 instances. 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 \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}(0.35−−0.50 μm/px)\end{document} against Faster R-CNN, YOLOv11, and Deformable DETR using COCO-style metric under the Olar et al. evaluation protocol, bootstrap-resampled FROC analysis, and targeted stratifications for local contrast and pair-separation recall. Results: Our method improved overall Average Precision (AP) across the vast majority of test cohorts, including a nearly 19% relative increase in high-resolution ALDH1L1 images. Crucially, stratified evaluations revealed significant improvements in localizing low-contrast targets and resolving densely crowded cell pairs without relying on heuristic filtering. FROC analysis confirmed superior sensitivity at clinically relevant low false-positive rates. Conclusions: Integrating spatial priors with global reasoning effectively mitigates the sensitivity-precision trade-off. This yields a robust detector for astrocyte soma localization resilient to stain and resolution shifts, serving as a highly reliable upstream component for quantitative digital pathology workflows. Supplementary Information: The online version contains supplementary material available at https://doi.org/10.1186/s12880-026-02678-5.
SN - 1471-2342
ER -