Cervical OCT image classification using contrastive masked autoencoders with Swin Transformer.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oncology

Background:

• Cervical cancer is a global health threat.
• Optical coherence tomography (OCT) shows promise for non-invasive diagnosis.
• Lack of labeled cervical OCT data limits deep learning applications.

Purpose of the Study:

• Propose CMSwin, a self-supervised learning (SSL) framework for cervical OCT images.
• Utilize abundant unlabeled cervical OCT data.
• Overcome data scarcity for deep learning models.

Main Methods:

• Developed a novel self-supervised learning framework (CMSwin).
• Combined masked image modeling (MIM) with contrastive learning.
• Employed Swin-Transformer architecture with mixed image encoding and contrastive losses.

Main Results:

• CMSwin outperformed state-of-the-art SSL approaches on a large multi-center dataset.
• Achieved performance comparable to or exceeding skilled medical experts in identifying high-risk lesions.
• Validated on multi-center and external datasets from top Chinese hospitals.

Conclusions:

• CMSwin has significant potential to aid gynecologists in interpreting cervical OCT images.
• The integrated GradCAM module offers visualization and interpretability for efficient diagnosis.
• Facilitates intelligent interpretation of cervical OCT images in clinical settings.