Deepfake Histologic Images for Enhancing Digital Pathology.

Area of Science:

• Digital pathology
• Computational pathology
• Artificial intelligence in medicine

Background:

• Histopathology relies on expert microscopic examination of tissue slides.
• Deep learning models show promise in tissue diagnostics but require extensive annotated data.
• Generating synthetic histology images can overcome data limitations and costs.

Purpose of the Study:

• To develop a generative adversarial network (GAN) for synthesizing realistic histology images.
• To enable the generation of both common and rare disease morphologies.
• To evaluate the utility of synthetic images in augmenting diagnostic capabilities.

Main Methods:

• Developed a class-conditional GAN to synthesize pathology images.
• Trained the GAN on prostate and colon tissue datasets.
• Assessed synthetic image realism and diagnostic utility with machine learning models and pathologist evaluation.

Main Results:

• Synthetic data performed comparably to real data in training diagnostic deep learning models.
• Pathologists could not differentiate between real and synthetic images.
• Similar interobserver agreement for prostate cancer grading was observed for real and synthetic data.
• The GAN successfully reproduced complex microenvironment morphologies in colon biopsies.

Conclusions:

• The proposed GAN framework effectively generates realistic histology images with diagnostic relevance.
• Synthetic images can augment machine learning diagnostic models and assist pathologists.
• The approach allows for the creation of novel and rare morphologies for research and training.