Semi-supervised learning for colposcopic image classification using generative adversarial networks

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computational Pathology

Background:

• Traditional colposcopy screening is limited by physician expertise and diagnostic accuracy.
• Uneven distribution of healthcare resources impacts access to expert colposcopic interpretation.
• Developing AI-assisted tools can address these limitations in cervical cancer screening.

Purpose of the Study:

• To develop an AI-assisted framework for generating and interpreting colposcopic images.
• To improve the accuracy and reliability of colposcopic image analysis using machine learning.

Main Methods:

• Proposed a Feature-pyramid and Squeeze-excitation Generative Adversarial Network (FSGAN) for realistic image generation.
• Developed a CenSwin Transformer model for region-focused colposcopic image classification.
• Utilized FSGAN-generated images and a dynamic pseudo-label semi-supervised learning strategy for model training.

Main Results:

• CenSwin Transformer achieved 90.39% accuracy and 86.10% recall in 12 experiments.
• Demonstrated statistically significant improvements in accuracy and F1-score compared to Swin Transformer.
• Validated on an independent dataset, achieving 85.26% accuracy and 71.11% recall.

Conclusions:

• The FSGAN-assisted semi-supervised learning framework and CenSwin Transformer enhance colposcopic image recognition.
• The AI model shows potential as an auxiliary tool for colposcopic image interpretation.
• This technology can support primary healthcare settings and areas with limited access to specialists.