Deep contrastive learning based tissue clustering for annotation-free histopathology image analysis.

Area of Science:

• Digital pathology
• Computational biology
• Machine learning in healthcare

Background:

• Supervised deep convolutional neural networks (CNNs) show promise in whole slide images (WSIs) processing.
• Manual annotation of gigapixel WSIs is labor-intensive and error-prone, creating a bottleneck for developing WSI diagnosis models.

Purpose of the Study:

• To develop a deep learning-based, self-supervised histopathology image analysis workflow for tissue classification without annotations.
• To overcome the limitations of manual annotation in digital pathology.

Main Methods:

• Utilized a self-supervised training scheme inspired by contrastive learning for annotation-free WSI patch analysis.
• Developed a multi-scale encoder network for extracting pathology-specific contextual features.
• Employed a silhouette coefficient-based recursive scheme for refining tissue clusters and building a tissue dictionary for cancer diagnosis.

Main Results:

• The method achieved competitive accuracy in identifying different tissues without annotation (0.9364/0.9325 in colorectal/sentinel lymph WSIs).
• Outperformed other unsupervised baseline methods in tissue identification.
• Demonstrated high performance in distinguishing benign/malignant polyps with an AUC score of 0.99 in a clinical cohort.

Conclusions:

• The proposed deep contrastive learning method enables learning from raw WSIs for tissue differentiation without annotation.
• The method shows potential as a quantitative and qualitative tool to assist pathologists in disease diagnosis.
• Tested across three datasets, demonstrating its versatility and applicability in digital pathology.