Deep Cervix Model Development from Heterogeneous and Partially Labeled Image Datasets.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computational Biology

Background:

• Cervical cancer is a major global health concern for women.
• Early detection of cervical precancer is crucial but limited by expert scarcity and interpretation variability.
• Automated cervical image classification systems are needed to assist experts.

Purpose of the Study:

• To develop a robust pretrained cervix model using heterogeneous and partially labeled cervical image datasets.
• To explore the application of self-supervised learning (SSL) and federated self-supervised learning (FSSL) for cervical image analysis.
• To improve the accuracy and reliability of cervical precancer detection through automated systems.

Main Methods:

• Utilized self-supervised learning (SSL) to pretrain a cervix model on partially labeled cervical image datasets.
• Implemented federated self-supervised learning (FSSL) to train the model without direct data sharing, addressing privacy concerns.
• Fine-tuned the pretrained cervix model for task-specific classification on two distinct datasets with varying labeling criteria.

Main Results:

• The dataset-specific SSL-pretrained cervix model demonstrated a 2.5% increase in classification accuracy compared to an ImageNet pretrained model.
• Combining images from both datasets for SSL further boosted classification accuracy by an additional 1.5%.
• Federated self-supervised learning (FSSL) showed superior performance compared to the dataset-specific SSL model.

Conclusions:

• SSL and FSSL are effective methods for developing high-performance cervix models from heterogeneous, partially labeled datasets.
• FSSL offers a viable solution for collaborative model development while respecting data privacy and sharing restrictions.
• The developed automated system has the potential to augment expert capabilities in cervical precancer screening and diagnosis.