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

Introduction: Traditional colposcopy screening faces several challenges, including heavy reliance on physician expertise, relatively high misdiagnosis rates, and uneven distribution of primary healthcare resources. To address these limitations, this study aimed to develop an artificial intelligence-assisted framework for colposcopic image generation and interpretation. Methods: We proposed a Feature-pyramid and Squeeze-excitation Generative Adversarial Network (FSGAN) to generate realistic colposcopic images and a region-focused recognition model named CenSwin Transformer for image classification. FSGAN-generated images were incorporated into model training, and a dynamic pseudo-label semi-supervised learning strategy was applied to improve recognition performance. The proposed model was evaluated through 12 repeated experiments and further validated on an additional independent dataset. Results: Across 12 repeated experiments, CenSwin Transformer achieved an accuracy of 90.39 ± 0.67%, recall of 86.10 ± 4.33%, and F1-score of 0.8645 ± 0.0111. Compared with Swin Transformer, CenSwin Transformer showed statistically significant improvements in accuracy and F1-score. On the independent validation dataset, the model achieved an accuracy of 85.26% and recall of 71.11%. Discussion: The proposed FSGAN-assisted semi-supervised learning framework and CenSwin Transformer model improved colposcopic image recognition performance, particularly in accuracy and F1-score. These findings suggest that the model may serve as an auxiliary second-opinion tool for colposcopic image interpretation, especially in primary healthcare settings or regions with limited access to experienced colposcopists.