Alleviating tiling effect by random walk sliding window in high-resolution histological whole slide image synthesis

Area of Science:

• Computational pathology
• Digital pathology
• Biomedical imaging

Background:

• Multiplex immunofluorescence (MxIF) allows simultaneous detection of over 20 molecular markers on a single tissue section.
• Hematoxylin and eosin (H&E) staining is standard for histology but typically incompatible with MxIF on the same section.
• Virtual H&E staining from MxIF using deep learning faces challenges with tiling artifacts in whole slide image (WSI) synthesis.

Purpose of the Study:

• To develop a deep learning-based method for unpaired, high-resolution virtual H&E WSI synthesis from MxIF WSIs.
• To mitigate tiling effects commonly encountered in cross-stain image synthesis for digital pathology.
• To enhance the utility of MxIF data by enabling simultaneous H&E visualization.

Main Methods:

• Extended the CycleGAN framework with simultaneous nuclei and mucin segmentation as spatial constraints.
• Implemented a random walk sliding window shifting strategy during inference to reduce tiling artifacts.
• Developed a method for synthesizing virtual H&E WSIs from MxIF WSIs with 27 markers.

Main Results:

• Achieved a 56% performance gain in downstream cell segmentation tasks through spatially constrained synthesis.
• Reduced tiling effects in synthesized H&E WSIs.
• Decreased computational resource usage by 50% during inference without performance compromise.

Conclusions:

• The proposed spatially constrained deep learning method effectively generates virtual H&E WSIs from MxIF data, overcoming tiling artifacts.
• The random walk sliding window inference strategy offers an efficient and generalizable solution for high-resolution WSI synthesis.
• This approach enhances the integration of multi-modal imaging data in digital pathology research.