Contrastive Multiple Instance Learning: An Unsupervised Framework for Learning Slide-Level Representations of Whole...

Area of Science:

• Computational pathology
• Artificial intelligence in medicine
• Digital pathology

Background:

• Current computational pathology methods often rely on semi- or weakly-supervised approaches needing slide-level labels.
• The absence or irrelevance of slide-level labels, particularly in clinical trials, necessitates annotation-free methods.
• Whole-slide images (WSIs) contain rich information that can be leveraged without explicit labels.

Purpose of the Study:

• To develop a fully unsupervised method for learning meaningful and compact representations from WSIs.
• To enable the analysis of unlabeled WSIs for various computational pathology tasks.
• To create a self-supervised learning framework for extracting slide-level features.

Main Methods:

• A tile-wise encoder is trained using SimCLR for initial feature extraction.
• Tile embeddings are fused using an attention-based multiple-instance learning framework to generate slide-level representations.
• Contrastive loss is employed to attract and repel intra-slide and inter-slide embeddings, facilitating self-supervised learning.

Main Results:

• The method achieved an AUC of 0.8641 ± 0.0115 for non-small cell lung cancer (NSCLC) subtyping.
• It obtained a correlation (R^2) of 0.5740 ± 0.0970 for breast cancer proliferation scoring (TUPAC16).
• Ablation studies confirmed that the learned unsupervised feature space can be effectively fine-tuned with small labeled datasets.

Conclusions:

• The proposed unsupervised method effectively learns representations from WSIs without requiring any annotations.
• This approach offers a novel way to benefit from completely unlabeled WSIs in computational pathology.
• The method has the potential to significantly advance the analysis of digital pathology data, especially in scenarios with limited or no labels.