Cells Grouping Detection and Confusing Labels Correction on Cervical Pathology Images.

Area of Science:

• Digital Pathology
• Computational Medicine
• Oncology

Background:

• Cervical cancer is a leading health threat for women globally.
• Early detection through screening is crucial for prevention and treatment.
• Automated pathological image analysis offers potential to enhance diagnostic efficiency and accuracy.

Purpose of the Study:

• To develop an advanced cervical cell detection network, PGCC-Net.
• To leverage clinical prior knowledge and address challenges of ambiguous cell labeling in deep learning models.
• To improve the accuracy and efficiency of cervical precancerous lesion detection.

Main Methods:

• Proposed PGCC-Net, a cervical cell detection network incorporating prior knowledge.
• Implemented cell grouping detection using clinical prior knowledge to learn cell structures.
• Developed a label correction module utilizing feature similarity and feature centers to resolve ambiguous cell annotations.
• Validated the model on public and private datasets.

Main Results:

• PGCC-Net demonstrated superior performance compared to state-of-the-art cervical cell detection methods.
• The model effectively utilized clinical prior knowledge for cell grouping and refined detection.
• The label correction module successfully addressed challenges posed by ambiguous cell classifications.
• Experimental validation confirmed the model's effectiveness on large datasets.

Conclusions:

• PGCC-Net offers a significant advancement in automated cervical cell detection.
• The integration of clinical prior knowledge and label correction enhances deep learning model performance.
• This approach holds promise for improving cervical cancer screening and diagnosis.