MUST: Multi-style virtual staining with incomplete pairs.

Area of Science:

• Digital pathology
• Computational imaging
• Artificial intelligence in medicine

Background:

• Multi-style virtual staining enhances histological image analysis by generating diverse staining modalities.
• Clinical adoption is hindered by incomplete paired training data due to tissue degradation and artifacts.
• Existing virtual staining methods require perfectly paired datasets, limiting real-world applicability.

Purpose of the Study:

• To introduce MUST (MUlti-style virtual STaining), a novel approach for virtual staining using incomplete supervision.
• To address the challenge of incompletely paired training data in multi-style virtual staining.
• To reformulate virtual staining as progressive cross-modality refinement.

Main Methods:

• MUST employs two synergistic components: Collaborative Denoising (CoDe) and Semantic Preservation (SP).
• CoDe utilizes cross-modality cross-attention within a latent diffusion model for information exchange under incomplete supervision.
• SP enhances cross-modal consistency via contrastive learning and generates pseudo-supervision from confident predictions.

Main Results:

• MUST significantly outperforms state-of-the-art methods on three histopathology datasets.
• The approach effectively mines cross-modality correlations despite incomplete data.
• High-confidence pseudo-supervision is generated from samples lacking ground truth.

Conclusions:

• MUST provides a robust solution for multi-style virtual staining with incomplete paired data.
• The method demonstrates improved clinical utility by overcoming data limitations.
• MUST facilitates advanced digital pathology applications through effective cross-modality learning.