Efficient Supervised Pretraining of Swin-Transformer for Virtual Staining of Microscopy Images

Area of Science:

• Computational biology
• Bioimaging
• Artificial intelligence in life sciences

Background:

• Fluorescence staining is crucial for cellular analysis but is time-consuming and limits simultaneous labeling.
• Virtual staining offers an alternative by eliminating chemical labeling, but deep learning models require extensive pretraining.
• Existing virtual staining methods often use private datasets and varied metrics, hindering fair comparison.

Purpose of the Study:

• To develop an efficient virtual staining method that reduces reliance on large datasets and computation.
• To introduce a novel pretraining strategy for transformer-based virtual staining models.
• To establish a standardized benchmark for evaluating virtual staining techniques.

Main Methods:

• Constructed a Swin-transformer model for virtual staining.
• Proposed an efficient supervised pretraining method using masked autoencoder (MAE) with downsampling and grid sampling to mask 75% of pixels.
• Developed a supervised proxy task to predict multi-style stained images instead of masked pixels.
• Created a standard benchmark using three public datasets for fair evaluation.

Main Results:

• The proposed pretraining method reduced pretraining time by 16 times compared to the original MAE.
• Achieved state-of-the-art performance on three benchmark datasets, both quantitatively and qualitatively.
• Ablation studies confirmed the effectiveness of the proposed pretraining strategy.

Conclusions:

• The developed efficient supervised pretraining method significantly enhances virtual staining performance while reducing computational demands.
• The standardized benchmark and baseline provide a valuable resource for future research in virtual staining.
• The approach offers a more accessible and effective solution for cellular constituent labeling in life sciences.