Foreground-aware Virtual Staining for Accurate 3D Cell Morphological Profiling

Area of Science:

• Cellular and Molecular Imaging
• Biomedical Engineering
• Machine Learning in Biology

Background:

• Transmitted-light microscopy offers low-cost, minimally invasive 3D imaging of cell morphology.
• Fluorescence microscopy provides high specificity and contrast but can be invasive.
• Virtual staining aims to combine the benefits of both by predicting fluorescence from label-free images using machine learning.

Purpose of the Study:

• To develop a novel virtual staining method, Spotlight, that overcomes limitations of existing approaches.
• To guide machine learning models to focus on biologically relevant cellular structures, rather than background noise and artifacts.
• To improve the accuracy and utility of virtual staining for downstream biological analyses.

Main Methods:

• Spotlight employs a machine learning approach for virtual staining.
• It utilizes histogram-based foreground estimation to mask pixel-wise loss calculations.
• A Dice loss on soft-thresholded predictions is used for shape-aware learning.

Main Results:

• Spotlight effectively guides the model to focus on significant cellular structures.
• The method improves the morphological representation of cells in 3D virtual stains.
• Pixel-level accuracy is preserved, leading to enhanced virtual stains.

Conclusions:

• Spotlight represents a significant advancement in virtual staining techniques for microscopy.
• The improved virtual stains generated by Spotlight are better suited for tasks like cell segmentation and profiling.
• This method enhances the value of label-free imaging for detailed cellular analysis.