MITIGATING OVER-SATURATED FLUORESCENCE IMAGES THROUGH A SEMI-SUPERVISED GENERATIVE ADVERSARIAL NETWORK.

Area of Science:

• Biomedical imaging
• Computational pathology
• Cellular and molecular imaging

Background:

• Multiplex immunofluorescence (MxIF) imaging provides spatial context for cell analysis.
• Saturation artifacts in MxIF images limit single-cell resolution and analysis.
• Existing correction methods often fail due to non-uniform saturation patterns.

Purpose of the Study:

• To develop a novel data-driven method for correcting multi-round saturation artifacts in MxIF images.
• To improve the accuracy of single-cell analysis in biomedical research by enhancing image quality.
• To introduce a hybrid generative adversarial network (HDmixGAN) for artifact correction.

Main Methods:

• A two-stage, high-resolution hybrid generative adversarial network (HDmixGAN) combining CycleGAN and pix2pixHD architectures was developed.
• CycleGAN was used to generate pseudo-paired data from unpaired over-saturated MxIF datasets.
• Pix2pixGAN was trained on both real paired and synthetic data from multiple DAPI staining rounds.

Main Results:

• The HDmixGAN method demonstrated improved performance in correcting saturation artifacts.
• Validation in a downstream nuclei detection task showed a 6% increase in F1 score compared to baseline methods.
• The approach effectively addresses multi-round saturation challenges in MxIF imaging.

Conclusions:

• The proposed HDmixGAN method offers a specialized solution for multi-round saturation artifacts in MxIF.
• This technique enhances cell analysis accuracy by improving the quality of biomedical images.
• The developed method represents a significant advancement in addressing image quality challenges in MxIF research.