Automated Screening of Precancerous Cervical Cells Through Contrastive Self-Supervised Learning.

Area of Science:

• Computational pathology
• Medical imaging analysis
• Artificial intelligence in healthcare

Background:

• Cervical cancer screening relies on labor-intensive cytology, requiring expert analysis of numerous cells.
• Current methods face challenges in efficiency and expert dependency for detecting precancerous lesions.

Purpose of the Study:

• To develop a novel distribution-augmented, contrastive self-supervised learning approach for detecting abnormal squamous cervical cells.
• To enhance the differentiation between normal and high-grade precancerous cervical cells (HSILs, ASC-H) using cytological images.

Main Methods:

• Utilized contrastive self-supervised learning with color augmentations on cytological images.
• Trained the model exclusively on normal cervical cell images.
• Employed kernel density estimation (KDE) to analyze cell type distributions and identify abnormalities.

Main Results:

• Achieved high diagnostic accuracy in detecting abnormal cervical cells.
• Demonstrated robustness against color distribution shifts in images.
• Successfully identified high-grade squamous intraepithelial lesions (HSILs) and atypical squamous cells-cannot exclude HSIL (ASC-H).

Conclusions:

• The proposed AI approach significantly improves cervical cancer screening accuracy.
• The method reduces the workload for cytopathologists, leading to more efficient screening programs.
• This contributes to advancing early detection strategies for cervical cancer.