Reproduce “Staining normalization in histopathology: Method benchmarking using multicenter dataset” on OpenResearch

About this paper

Umair Khan, Jouni Härkönen, Marjukka Friman, Leena Latonen, Teijo Kuopio, Pekka Ruusuvuori

Researchers from the University of Turku and collaborators created a unique multicenter H&E stained slide dataset from 66 laboratories across 11 countries, utilizing identical tissue blocks to isolate staining variability. This dataset was used to benchmark various traditional and deep learning-based stain normalization methods, revealing that traditional methods like Histogram Matching often achieved comparable or superior color transfer accuracy to GAN-based approaches, while deep learning methods sometimes introduced visual artifacts.

Key insights

• No single stain normalization method proved universally superior across all metrics and types of staining issues, indicating the need for context-specific method selection.
• Traditional methods, particularly Histogram Matching and Reinhard, often achieved performance comparable to or better than deep learning GANs in terms of color transfer accuracy and qualitative resemblance to the reference.
• Deep learning GANs, while providing reasonable results, showed a propensity for generating generative artifacts and did not conclusively outperform traditional methods in this specific content-consistent, single-WSI-per-lab setup.

Summary by alphaXiv.