Knowledge-Based Representation Learning for Nucleus Instance Classification From Histopathological Images.

Area of Science:

• Digital Pathology
• Computational Biology
• Medical Image Analysis

Background:

• Nucleus classification in H&E-stained histopathology images is crucial for quantitative digital pathology analysis.
• Current multi-class classification methods are limited by annotation scale and often neglect contextual information.
• Effective nucleus classification requires incorporating contextual details surrounding individual cell instances.

Purpose of the Study:

• To develop a novel approach for nucleus classification in digital pathology that leverages contextual information.
• To design a structured input format combining image patches and instance masks for enhanced classification.
• To pre-train a feature extraction model using a large-scale unlabeled dataset to reduce annotation burden.

Main Methods:

• A structured input format comprising a content-rich image patch and a target instance mask was designed.
• A Structured Triplet representation learning framework with customized sampling strategies was proposed for nucleus instances.
• Pre-training was performed on a large-scale unlabeled dataset of H&E-stained pathology images.
• Auxiliary attribute learning and self-supervised learning branches were incorporated to boost performance.

Main Results:

• The proposed framework significantly improves nucleus classification accuracy compared to state-of-the-art methods.
• Pre-training on the new dataset with the Structured Triplet framework reduces the need for extensive labeling.
• The model demonstrates enhanced performance due to the structured input and auxiliary branches.

Conclusions:

• The developed structured input and Structured Triplet framework offer a powerful solution for nucleus classification in digital pathology.
• This approach effectively utilizes contextual information, leading to improved accuracy and reduced annotation requirements.
• The release of a new large-scale dataset will facilitate further research in this domain.