Artificial intelligence workflow quantifying muscle features on Hematoxylin–Eosin stained sections reveals dystrophic phenotype amelioration upon treatment

Abstract

Abstract Cell segmentation is a key step for a wide variety of biological investigations, especially in the context of muscle science. Currently, automated methods still struggle to perform skeletal muscle fiber quantification on Hematoxylin-Eosin (HE) stained histopathological whole slide images due to low contrast. On the other hand, the Deep Learning algorithm Cellpose offers new perspectives considering its increasing adoption for segmentation of a wide range of cells. Combining two open-source tools, Cellpose and QuPath, we developed MyoSOTHES, an automated Myofibers Segmentation wOrkflow Tuned for HE Staining. MyoSOTHES enables solving segmentation inconsistencies encountered by default Cellpose model in presence of large range size cells and provides information related to muscle Feret’s diameter distribution and Centrally Nucleated Fibers, thus depicting muscle health and treatment effects. MyoSOTHES achieves high quality segmentation compared to baseline workflow with a detection F1-score increasing from 0.801 to 0.919 and a Root Mean Square Error (RMSE) on diameter improved by 31%. MyoSOTHES was validated on an animal study featuring gene transfer in $$\gamma$$ γ -Sarcoglycanopathy, for which dose-response effect is visible and conclusions drawn are consistent with those previously published. MyoSOTHES thus paves the way for wide quantification of HE stained muscle sections and retrospective analysis of HE labeled slices used in laboratories for decades.