RetCCL: Clustering-guided contrastive learning for whole-slide image retrieval.

Area of Science:

• Digital Pathology
• Computational Biology
• Medical Imaging Analysis

Background:

• Digitized whole-slide images (WSIs) enable computer-aided diagnosis, but content-based retrieval is challenging due to large image sizes.
• Accurate retrieval of similar WSIs is crucial for clinical diagnosis, research, and education.

Purpose of the Study:

• To develop a robust and accurate framework for whole-slide image (WSI)-level retrieval.
• To address challenges in encoding semantic content and measuring similarity in gigapixel histopathological images.
• To provide interpretable retrieval results by highlighting similar image regions.

Main Methods:

• Proposed a Retrieval with Clustering-guided Contrastive Learning (RetCCL) framework.
• Integrated a novel self-supervised feature learning method using unlabeled histopathological data.
• Employed a global ranking and aggregation algorithm for improved performance.

Main Results:

• The RetCCL framework achieved significant performance improvements on anatomical site and cancer subtype retrieval tasks using over 22,000 slides.
• Demonstrated around 10% improvement in average mMV@10 for anatomic site retrieval compared to state-of-the-art methods.
• Achieved a 24% performance improvement in patch retrieval on the TissueNet dataset using learned features.

Conclusions:

• The proposed RetCCL framework offers a robust and accurate solution for WSI-level retrieval.
• Self-supervised feature learning with unlabeled data enhances WSI retrieval without fine-tuning.
• The framework provides interpretable results, aiding pathologists in understanding retrieval outcomes.