An Immunofluorescence-Guided Segmentation Model in Hematoxylin and Eosin Images Is Enabled by Tissue Artifact...

Area of Science:

• Computational pathology
• Digital pathology
• Machine learning in histopathology

Background:

• Supervised deep learning models for histopathology require extensive, accurate ground truth annotations, which are time-consuming and costly to generate.
• Immunofluorescence (IF) staining offers molecular annotation but creates discrepancies between standard Hematoxylin and Eosin (H&E) and post-IF H&E (terminal H&E) stains, limiting its use for training.
• These discrepancies hinder the direct application of IF-derived annotations to conventional H&E images for training computer vision models.

Purpose of the Study:

• To develop a method for training accurate computer vision models on conventional H&E images without manual annotations.
• To overcome the staining discrepancies between conventional H&E and terminal H&E caused by IF processing.
• To enable the use of molecularly annotated data (via IF) for large-scale ground truth generation in digital pathology.

Main Methods:

• Utilized a cycle-consistent generative adversarial network (CycleGAN) to synthesize conventional H&E images that emulate the appearance of terminal H&E stains.
• Trained a deep learning segmentation model on these synthetically generated terminal H&E images for epithelium segmentation.
• Validated the segmentation model using IF staining of epithelial markers (cytokeratins) on the actual tissue.
• Integrated the CycleGAN stain transfer model with the segmentation model for performative epithelium segmentation on conventional H&E images.

Main Results:

• Synthetic image generation successfully emulated terminal H&E appearance in conventional H&E images.
• A deep learning model trained on synthetic data achieved accurate epithelium segmentation, validated by IF staining.
• The combined approach enabled performative epithelium segmentation directly on conventional H&E images.
• Demonstrated the feasibility of training segmentation models without human expert annotations by leveraging IF and generative models.

Conclusions:

• Generative adversarial networks can bridge the gap between conventional H&E and terminal H&E, enabling the use of molecular annotations for training computer vision models.
• This approach significantly accelerates the generation of accurate ground truth for digital pathology applications.
• It allows for the development of robust computer vision tools for histopathology using readily available conventional H&E data, bypassing the need for manual annotation.