Related Experiment Video
Updated: Aug 5, 2026
08:40
Quantitation of Protein Expression and Co-localization Using Multiplexed Immuno-histochemical Staining and Multispectral Imaging
Published on: April 8, 2016
Object detection in histology: A multi-dataset benchmark and test-time inference
Dragoș-Vasile Leordean1,2, Eugen-Richard Ardelean3
1Molecular Neuroscience Laboratory, Transylvanian Institute of Neuroscience, Cluj-Napoca, Romania.
Plos One
|July 29, 2026
Summary
Object detection models like YOLOv11 excel in analyzing histological images for medical diagnosis. However, all models struggle with small objects, highlighting areas for future research in automated medical image analysis.
Area of Science:
- Computer Vision
- Medical Image Analysis
- Histopathology
Background:
- Automated medical diagnosis relies heavily on medical image analysis.
- Object detection models are crucial for identifying pathological structures in images.
- Deep learning advancements have significantly impacted medical image analysis.
Purpose of the Study:
- To conduct a comprehensive comparative analysis of the latest object detection models for histological image analysis.
- To evaluate YOLO architectures (YOLOv8-YOLOv12) against alternatives like RT-DETR, YOLO-World, and YOLOE.
- To propose and evaluate Test-time Graph Similarity Propagation (TGSP) for refining object detection in histology.
Main Methods:
- Comparative evaluation of object detection models (YOLOv8-YOLOv12, RT-DETR, YOLO-World, YOLOE) on five histology datasets (BCNB, Nuclei, TNBC, MoNuSAC, CryoNuSeg).
- Standardized training protocols, consistent hyperparameters, and data augmentation were used.
- Performance was assessed using multiple metrics, inference time, and computational cost. TGSP was introduced as a test-time refinement method.
Main Results:
- YOLOv11 demonstrated strong performance across datasets, while YOLOv12 showed promise with new attention mechanisms.
- RT-DETR performed poorly on small objects, common in histology. All models struggled with small object detection.
- TGSP consistently matched or improved F1 scores compared to standard thresholds and TSBP, especially with poor base-model calibration.
Conclusions:
- YOLOv11 is a top performer for object detection in histology, but small object detection remains a challenge for all models.
- Failure cases include background false positives and missed detections, indicating areas for improvement.
- TGSP offers a scalable and effective method for enhancing object detection accuracy in histopathology, particularly when base model confidence is low.