[SEPTEMBER 14TH, 2023]
Authors: Guillaume Balezo , Christof A. Bertram , Cyprien Tilmant , Stéphanie Petit , Saima Ben Hadj , Rutger H.J. Fick
Topic: Digestive Diseases Pathology - Liver/Pancreas
Background & objectives
To highlight fibrotic collagen in liver disease, special stains like Masson's Trichrome (MT), Sirius Red (SR), or Saffron (HES) are required. This study aims to highlight collagen directly from Hematoxylin and Eosin (HE) stained liver biopsies using deep learning.
Methods
We obtained 11 retrospective liver cases with varying degrees of fibrosis. For each case, we collected three consecutive slides, each stained with HE, and separately overstained with either HES, MT, or SR. We predict the collagen on the HE (c-HE) using deep learning. We validate our approach by having two pathologists do METAVIR scoring on the special stains and c-HE.
Results
Results reveal a strong pixel-wise correlation between c-HE and each registered special stain (HES: 0.87, MT: 0.78, SR: 0.71) with p-value<1e-5, indicating accurate collagen content inference from HE-stained liver biopsies. Pathologist agreement on METAVIR scores was high, with P1 at 10/11 and P2 at 11/11 cases for both special stains and c-HE. This supports the potential for deep learning to replace special stains in detecting fibrosis, maintaining highquality METAVIR evaluations.
Conclusion
In conclusion, our study demonstrates that deep learning can accurately infer collagen content from HE-stained liver biopsies, both quantitatively and qualitatively, showing a high correlation with the standard special stains traditionally used to highlight collagen. Moreover, the agreement between pathologists on METAVIR scores remains high when using our digital HE-predicted collagen method, suggesting that this approach can potentially replace the need for special stains in detecting fibrosis without compromising the quality of METAVIR evaluations.
Download the e-poster
