SGCL: Spatial guided contrastive learning on whole-slide pathological images
Tiancheng Lin1, Zhimiao Yu1, Zengchao Xu2
1Shanghai Key Lab of Digital Media Processing and Transmission, Shanghai Jiao Tong University, China; MoE Key Lab of Artificial Intelligence, AI Institute, Shanghai Jiao Tong University, China.
Medical Image Analysis
|August 19, 2023
Summary
Spatial Guided Contrastive Learning (SGCL) enhances self-supervised learning for whole-slide pathological images (WSIs). SGCL leverages spatial information to improve performance on diverse downstream tasks, outperforming existing methods.
Area of Science:
- Computational pathology
- Machine learning
- Medical image analysis
Background:
- Self-supervised learning (SSL) excels in natural image analysis but underperforms on whole-slide pathological images (WSIs) due to their gigapixel resolution and complex patch structures.
- Existing SSL methods for natural images are suboptimal for WSIs, necessitating tailored approaches.
Purpose of the Study:
- To develop a novel self-supervised learning scheme, Spatial Guided Contrastive Learning (SGCL), specifically designed for WSIs.
- To leverage the inherent spatial properties of WSIs for more stable and effective self-supervision.
Main Methods:
- SGCL utilizes spatial proximity and multi-object priors for robust self-supervision.
- It expands intra- and inter-WSI invariance using spatial relationships and introduces spatial-guided multi-cropping for intra-patch invariance.
- A novel, theoretically validated loss function adaptively explores spatial information without supervision.
Main Results:
- SGCL achieves significant performance improvements over state-of-the-art pre-training methods on various downstream tasks and datasets.
- Ablation studies and visualizations confirm the effectiveness and aid in understanding the SGCL scheme.
Conclusions:
- SGCL offers a powerful new approach for self-supervised representation learning in whole-slide pathology.
- The method demonstrates the potential of exploiting WSI-specific spatial characteristics for improved machine learning performance in digital pathology.
