CS-CO: A Hybrid Self-Supervised Visual Representation Learning Method for H&E-stained Histopathological Images.

Area of Science:

• Computational histopathology
• Deep learning
• Medical image analysis

Background:

• Visual representation extraction is crucial in computational histopathology.
• Self-supervised learning (SSL) is promising for unlabeled histopathological images due to deep learning's power and annotation scarcity.
• Existing SSL methods for histopathology have limitations in versatility and representation capacity.

Purpose of the Study:

• To propose CS-CO, a hybrid SSL method for H&E-stained histopathological images.
• To integrate generative and discriminative approaches for robust visual representation learning.
• To improve the versatility and representation capacity of SSL in computational histopathology.

Main Methods:

• CS-CO employs a two-stage SSL approach: cross-stain prediction (CS) and contrastive learning (CO).
• A novel stain vector perturbation technique is introduced for data augmentation to aid contrastive learning.
• The method leverages domain-specific knowledge without requiring side information, ensuring rationality and versatility.

Main Results:

• CS-CO demonstrates effectiveness and robustness on patch-level tissue classification and slide-level cancer prognosis/subtyping tasks.
• Evaluations were conducted on three H&E-stained histopathological image datasets.
• Ablation studies confirmed that cross-staining prediction and contrastive learning components synergistically enhance performance.

Conclusions:

• CS-CO offers a versatile and effective solution for self-supervised visual representation learning in computational histopathology.
• The hybrid approach addresses limitations of previous SSL methods.
• The method shows strong performance on diverse downstream tasks, highlighting its practical utility.