While conventional fluorescence microscopy enables high-precision cellular observation, it poses a challenge in that it can damage living cells. In this study, we developed a system that combines non-invasive phase-contrast microscopy images with artificial intelligence (AI) to predict the location of mitochondria from unlabeled live-cell images. The output images generated by the trained model closely resembled actual fluorescence-stained images, and a certain level of reproducibility was confirmed. To improve the accuracy of the learning model, we introduced quantitative image evaluation metrics such as the Structural Similarity Index (SSIM) to complement traditional subjective evaluations. We also re-examined the preprocessing of stained images used as ground-truth labels, considering methods such as using the threshold that maximizes SSIM and using raw images without preprocessing. Furthermore, we introduced SSIM Loss as a new training loss function and compared its performance with Dice Loss. As a result, although some metrics such as SSIM and PSNR improved, there were cases where the actual accuracy of the output images decreased despite high evaluation scores. These results suggest that a comprehensive improvement of the system including the integration of multiple evaluation metrics, optimization of loss functions, and enhancement of preprocessing methods is a future challenge.
YAZAWA et al. (Wed,) studied this question.