Quantifying the recovery process of skeletal muscle on hematoxylin and eosin stained images via learning from label proportion.

Visual observing muscle tissue regeneration is used to measure experimental effect size in biological research to discover the mechanism of muscle strength decline due to illness or aging. Quantitative computer imaging analysis for support evaluating the recovery phase has not been established because of the localized nature of recovery and the difficulty in selecting image features for cells in regeneration. We constructed MyoRegenTrack for segmenting cells and classifying their regeneration phase in hematoxylin-eosin (HE) stained images. A straightforward approach to classification is supervised learning. However, obtaining detailed annotations for each fiber in a whole slide image is impractical in terms of cost and accuracy. Thus, we propose to learn individual recovery phase classification utilizing the proportions of cell class depending on the days after muscle injection to induce regeneration. We extract implicit multidimensional features from the HE-stained tissue images and train a classifier using weakly supervised learning, guided by their class proportion for elapsed time on recovery. We confirmed the effectiveness of MyoRegenTrack by comparing its results with expert annotations. A comparative study of the recovery relation between two different muscle injections shows that the analysis result using MyoRegenTrack is consistent with findings from previous studies.