Self-Supervised Triplet Contrastive Learning for Classifying Endometrial Histopathological Images.

Area of Science:

• Digital pathology
• Machine learning in medicine
• Oncology

Background:

• Accurate histopathological image analysis is vital for endometrial cancer diagnosis and treatment.
• Pathologist shortages hinder timely and precise diagnosis.
• Computer-aided diagnosis (CAD) systems, particularly deep learning (DL), offer automated solutions.

Purpose of the Study:

• To develop a novel self-supervised triplet contrastive learning (SSTCL) model for classifying endometrial histopathological images.
• To improve the accuracy and efficiency of endometrial disease diagnosis using limited annotated data.
• To reduce reliance on extensive human annotation for training diagnostic models.

Main Methods:

• Developed a self-supervised triplet contrastive learning (SSTCL) model with online and target branches.
• Incorporated a random mosaic masking (RMM) augmentation module for regularization.
• Utilized a bottleneck Transformer (BoT) model for self-attention and global information learning.
• Evaluated the model on public and in-house endometrial histopathological datasets.

Main Results:

• Achieved high four-class classification accuracies (up to 83.22%) on a public dataset with varying labeled data percentages.
• Obtained a 96.81% diagnostic accuracy on an in-house dataset.
• Outperformed existing state-of-the-art supervised and self-supervised methods on both datasets.

Conclusions:

• The developed SSTCL model demonstrates high accuracy and efficiency in diagnosing endometrial diseases from histopathological images.
• The model effectively utilizes limited human-annotated data, addressing a key challenge in digital pathology.
• This approach shows significant potential to assist pathologists in automated endometrial disease diagnosis.