Cluster-based histopathology phenotype representation learning by self-supervised multi-class-token hierarchical ViT.

Area of Science:

• Computational pathology
• Artificial intelligence in medicine
• Machine learning for medical imaging

Background:

• Clinical AI development requires extensive, expert-annotated datasets, increasing time and cost.
• Self-supervised learning (SSL) leverages unlabeled data to build domain-specific knowledge, improving AI model performance.
• Histopathology image analysis is crucial for disease diagnosis and research.

Purpose of the Study:

• Introduce CypherViT, a novel hierarchical Vision Transformer (ViT) for self-supervised representation learning in histopathology.
• Enable both coarse and fine-grained feature extraction from histopathological images.
• Develop a cost-effective and efficient method for clinical AI model training.

Main Methods:

• Developed CypherViT, a cluster-based, self-supervised, multi-class-token hierarchical ViT.
• Integrated CypherViT into the DINO SSL framework for training on unlabeled breast cancer histopathology images.
• Employed a hierarchical feature agglomerative attention module with multiple classification tokens for enhanced feature learning.

Main Results:

• CypherViT successfully learned semantically meaningful regions of interest aligned with morphological phenotypes.
• The trained model demonstrated generalizability as a feature extractor for colorectal cancer images.
• Achieved promising performance in patch-level tissue phenotyping across four public datasets, outperforming existing SSL and transfer learning methods.

Conclusions:

• CypherViT offers a robust and generalizable approach to histopathology representation learning using self-supervised methods.
• This method significantly reduces reliance on large, annotated datasets, lowering AI development costs.
• CypherViT shows substantial advantages over current state-of-the-art SSL and traditional transfer learning techniques in medical imaging.