Journal of Pathology Informatics (Jan 2022)

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

  • Mario Siller,
  • Lea Maria Stangassinger,
  • Christina Kreutzer,
  • Peter Boor,
  • Roman D. Bulow,
  • Theo J.F. Kraus,
  • Saskia von Stillfried,
  • Soraya Wolfl,
  • Sebastien Couillard-Despres,
  • Gertie Janneke Oostingh,
  • Anton Hittmair,
  • Michael Gadermayr

Journal volume & issue
Vol. 13
p. 100168

Abstract

Read online

Background: The fast acquisition process of frozen sections allows surgeons to wait for histological findings during the interventions to base intrasurgical decisions on the outcome of the histology. Compared with paraffin sections, however, the quality of frozen sections is often strongly reduced, leading to a lower diagnostic accuracy. Deep neural networks are capable of modifying specific characteristics of digital histological images. Particularly, generative adversarial networks proved to be effective tools to learn about translation between two modalities, based on two unconnected data sets only. The positive effects of such deep learning-based image optimization on computer-aided diagnosis have already been shown. However, since fully automated diagnosis is controversial, the application of enhanced images for visual clinical assessment is currently probably of even higher relevance. Methods: Three different deep learning-based generative adversarial networks were investigated. The methods were used to translate frozen sections into virtual paraffin sections. Overall, 40 frozen sections were processed. For training, 40 further paraffin sections were available. We investigated how pathologists assess the quality of the different image translation approaches and whether experts are able to distinguish between virtual and real digital pathology. Results: Pathologists' detection accuracy of virtual paraffin sections (from pairs consisting of a frozen and a paraffin section) was between 0.62 and 0.97. Overall, in 59% of images, the virtual section was assessed as more appropriate for a diagnosis. In 53% of images, the deep learning approach was preferred to conventional stain normalization (SN). Conclusion: Overall, expert assessment indicated slightly improved visual properties of converted images and a high similarity to real paraffin sections. The observed high variability showed clear differences in personal preferences.

Keywords