C-Norm boosts cervical cancer screening with cell distribution normalization
Outperforms mainstream detectors on real clinical ThinPrep cytology images
Cervical cancer screening via ThinPrep Cytologic Test (TCT) is critical but suffers from inconsistent manual diagnosis and poor AI performance in real conditions. The main issues are unbalanced spatial distribution of cells and a shortage of high-quality annotated data. To tackle this, researchers introduce Cell-Distribution Normalization (C-Norm), which decouples abnormal and normal cells from original TCT images and re-synthesizes them into a uniform distribution. This preprocessing step eliminates distribution bias that degrades model generalization.
C-Norm is built on a hybrid architecture combining YOLOv12’s robust detection capabilities with DINOv3’s superior feature representation. This synergy allows the model to detect subtle morphological nuances essential for accurate recognition of abnormal cells. Extensive experiments show that C-Norm significantly surpasses mainstream detection algorithms on real clinical datasets. The complete code is open-sourced, enabling replication and further research in automated cytopathology.
- Decouples abnormal and normal cells from TCT images to re-synthesize a uniform distribution, eliminating spatial bias.
- Integrates YOLOv12 (detection) with DINOv3 (feature extraction) for fine-grained morphological recognition.
- Achieves state-of-the-art results, outperforming standard detectors on real clinical ThinPrep cytology data.
Why It Matters
Standardizes AI-based cervical cancer screening, reducing diagnostic inconsistency and improving early detection rates in clinical practice.