Classification of breast cancer histopathology images using a modified supervised contrastive learning method

Area of Science:

• Medical Image Analysis
• Artificial Intelligence in Healthcare
• Computational Pathology

Background:

• Deep neural networks excel in medical image tasks but overfit with limited data.
• Overfitting in deep learning models hinders accurate disease classification and detection.
• Robustness in medical image analysis models is crucial for reliable diagnostic tools.

Purpose of the Study:

• To improve supervised contrastive learning for medical image analysis with limited data.
• To enhance model robustness against overfitting using domain-specific augmentations.
• To increase the accuracy of breast cancer detection from histopathology images.

Main Methods:

• Integrated self-supervised pre-training with a two-stage supervised contrastive learning strategy.
• Employed a modified supervised contrastive loss to reduce false negatives and eliminate false positives.
• Introduced a relaxing mechanism to refine positive and negative pairs based on similarity.

Main Results:

• Achieved a 1.45% increase in image-level classification accuracy on the BreakHis dataset.
• Reached an absolute accuracy of 93.63% in breast cancer histopathology image classification.
• Demonstrated improved model robustness and more appropriate representation learning.

Conclusions:

• The proposed method effectively enhances deep learning models for medical image analysis, particularly in low-data scenarios.
• Leveraging domain-specific augmentations and refined contrastive learning improves diagnostic accuracy.
• The approach offers a robust solution for breast cancer detection, paving the way for advanced computational pathology tools.