Fig. 2: CM recognition and differentiation efficiency evaluation by deep learning on bright-field images.

a Morphological characteristics of CMs and non-CMs on day 12 bright-field images and cTnT fluorescence results. Scale bar, 500 μm. b Schematic overview of deep learning-based cTnT fluorescence prediction from live-cell bright-field images at CM stage (stage III). The pix2pix model was trained with pairs of bright-field and fluorescence images. The trained model could predict the fluorescent labels for new bright-field images. For evaluation, the model’s prediction was further compared with the true cTnT fluorescence image. c Typical predicted results of CM recognition on bright-field image patches (containing cTnT⁺ CMs) from the test set. Each column from left to right represents: live-cell bright-field image patches; true cTnT fluorescence results; predicted cTnT fluorescence results. Scale bar, 250 μm. d Typical predicted results of CM recognition on whole-well bright-field images. Each column from left to right: true cTnT fluorescence result; predicted cTnT fluorescence result; the heatmap comparing the predicted with the true fluorescence pixel intensities, with Pearson’s r values provided. Scale bar, 1 mm. e Comparison of the true and predicted Differentiation Efficiency Indexes (i.e., total fluorescence intensity) for whole-well fluorescence images. True and predicted Differentiation Efficiency Indexes were normalized between 0 and 100%. n = 36 wells. f Correlation between the true and predicted Differentiation Efficiency Indexes, with Pearson’s r value. True and predicted Differentiation Efficiency Indexes were normalized between 0 and 100%. n = 36 wells.