A Key-Points Based Anchor-Free Cervical Cell Detector.

Area of Science:

• Medical Imaging
• Computer Vision
• Biomedical Engineering

Background:

• Cervical cell detection is vital for early cervical cancer screening.
• Current anchor-based methods (e.g., Faster R-CNN, YOLOv3) have limitations due to predefined hyperparameters and inflexibility with cell morphology.
• These limitations impact the accuracy of cell localization and classification.

Purpose of the Study:

• To develop an improved cervical cell detection system.
• To overcome the limitations of anchor-based methods by introducing an anchor-free strategy.
• To enhance the accuracy and robustness of cervical cell detection in cytological screening.

Main Methods:

• A key-points based anchor-free cervical cell detector was developed, building upon the YOLOv3 architecture.
• The method replaces predefined anchors with a key-points representation for initial cell prediction.
• Incorporated were PAFPN for enhanced feature hierarchy, GIoU loss for small cell optimization, alongside focal loss and smooth L1 loss.

Main Results:

• The proposed key-points based anchor-free method demonstrated effectiveness in cervical cell detection.
• The system showed improved cell localization and classification compared to conventional anchor-based approaches.
• The method proved robust across different liquid-based preparation styles (drop-slide, membrane-based, sedimentation).

Conclusions:

• The developed anchor-free detector offers a more desirable cell localization effect through refinement.
• This approach enhances the overall performance and adaptability of cervical cell detection systems.
• The findings support the potential of key-points based anchor-free methods for reliable cervical cancer screening.