On the acceptance of “fake” histopathology: A study on frozen sections optimized with deep learning

This research explores using deep learning to improve the quality of frozen section images, which are often lower quality than standard paraffin sections. The goal is to enhance these images to aid pathologists in making more accurate diagnoses during surgery. Generative adversarial networks (GANs) were used to translate frozen sections into virtual paraffin sections. Pathologists then assessed the quality of these translated images and tried to distinguish them from real paraffin sections. The study found that pathologists often preferred the deep learning-enhanced images for diagnosis and had difficulty distinguishing them from real paraffin sections, suggesting that this technology could improve diagnostic accuracy.

• 1
  Pathologists' detection accuracy of virtual paraffin sections was between 0.62 and 0.97, indicating a challenge in distinguishing real from AI-generated images.
• 2
  In 59% of images, the virtual section was assessed as more appropriate for a diagnosis than the original frozen section.
• 3
  The deep learning approach was preferred to conventional stain normalization in 53% of images, suggesting an advantage of AI-driven enhancement.