Fluorescent marker prediction for non-invasive optical imaging in bovine satellite cells using deep learning

Assessing the quality of bovine satellite cells (BSCs) is vital for advancing tissue engineered muscle constructs with applications in sustainable protein research. In this study, we present a non-invasive deep learning approach for optical imaging that predicts fluorescent markers directly from brightfield microscopy images of BSC cultures. Using a convolutional neural network based on the U-Net architecture, our method simultaneously predicts two key fluorescent signals, specifically DAPI and Pax7, which serve as biomarkers for cell abundance and differentiation status. An image preprocessing pipeline featuring fluorescent signal denoising was implemented to enhance prediction performance and consistency. A dataset comprising 48 biological replicates was evaluated using statistical metrics such as the Pearson <i>r</i> (correlation coefficient), the mean squared error (MSE), and the structural similarity Index (SSIM). For DAPI, denoising improved the Pearson <i>r</i> from 0.065 to 0.212 and SSIM from 0.047 to 0.761 (with MSE increasing from 9.507 to 41.571). For Pax7, the Pearson <i>r</i> increased from 0.020 to 0.124 and MSE decreased from 44.753 to 18.793, while SSIM remained low, reflecting inherent biological heterogeneity. Furthermore, enhanced visualization techniques, including color mapping and image overlay, improved the interpretability of the predicted outputs. These findings underscore the importance of optimized data preprocessing and demonstrate the potential of AI to advance non-invasive optical imaging for cellular quality assessment in tissue biology. This work also contributes to the broader integration of machine learning and computer vision methods in biological and agricultural applications.