Unsupervised Representation Learning for Tissue Segmentation in Histopathological Images: From Global to Local...

Area of Science:

• Computational pathology
• Deep learning
• Medical image analysis

Background:

• Tissue segmentation is crucial in computational pathology but hindered by annotation difficulties.
• Existing contrastive learning methods focus on global features, limiting pixel-level discrimination for segmentation tasks.

Purpose of the Study:

• To develop a contrastive learning framework that encodes multi-granularity features for tissue segmentation without annotations.
• To improve deep learning model performance in scenarios with limited or sparse annotations.

Main Methods:

• Designed three contrastive learning tasks: image-level (component discrimination), superpixel-level (prototype discrimination), and pixel-level (spatial smoothness).
• Employed a global-to-local pre-training strategy to capture fine-grained, domain-specific patterns.
• Validated on two tissue segmentation datasets under limited and sparse annotation conditions.

Main Results:

• The proposed framework effectively captures domain-specific and fine-grained patterns.
• Learned representations are transferable to various histopathological tissue segmentation tasks.
• Outperformed existing contrastive learning methods in experiments with limited/sparse annotations.

Conclusions:

• The global-to-local contrastive learning strategy enhances tissue segmentation by encoding multi-granularity features.
• This approach significantly improves performance in weakly supervised and semi-supervised settings.
• The framework offers a robust solution for annotation-scarce computational pathology challenges.