Generating region proposals for histopathological whole slide image retrieval

Area of Science:

• Digital Pathology
• Computational Imaging
• Machine Learning in Medicine

Background:

• Content-based image retrieval (CBIR) is crucial for histopathological image analysis.
• Analyzing whole slide images (WSIs) presents challenges in identifying relevant regions of interest (ROIs) for training.
• Pathologist annotations are scarce, making automated ROI generation and retrieval with limited labels a significant task.

Purpose of the Study:

• To develop an unsupervised method for generating accurate region proposals from histopathological WSIs.
• To introduce a novel semi-supervised hashing technique for efficient and precise image retrieval.
• To enable effective histopathological image analysis and retrieval with minimal labeled data.

Main Methods:

• Implemented an unsupervised region proposing method based on Selective Search for WSIs.
• Defined nucleus-oriented similarity measures and a Nucleus-Cytoplasm color space for region merging.
• Developed a semi-supervised hashing method using Latent Dirichlet Allocation for feature extraction and Supervised Hashing for binary code generation.

Main Results:

• The region proposing method generated 7.3 thousand contoured regions per WSI, aligning with 95.8% of pathologist-annotated ROIs.
• The hashing method achieved 91% precision in retrieving images from a database of 136,000 WSIs in 0.29 seconds, using only 10% labeled data.

Conclusions:

• Unsupervised region proposals can effectively predict lesions in histopathological WSIs.
• These region proposals can serve as training samples for machine learning models in image retrieval.
• The proposed hashing method enables fast and accurate retrieval with limited labels, with potential applications in computer-aided diagnosis systems.