Ultrastructural Morphometry of Mitochondria: Comparison Between Conventional Operator‐Dependent and Artificial Intelligence (<scp>AI</scp>)‐Operated Machine Learning Methods

Morphometric analysis of digital images is fundamental to substantiate the visual observations with objective quantitative data suitable for statistical analysis. The recent advances in artificial intelligence (AI) have allowed the development of machine learning (ML) protocols for automated morphometry. Transmission electron microscopy (TEM) morphometry requires that the ultrastructural details be recognized and interpreted by a trained observer; this makes adapting AI-operated protocols to TEM particularly challenging. In this study, we have checked the accuracy of the results of mitochondrial morphometry yielded by a ML method by comparison with those obtained manually by a trained observer on the same TEM micrographs (magnification ×50,000) of cultured cells with different energy metabolism (overall n = 26). The measured parameter was the ratio between the total length of the mitochondrial cristae and the corresponding mitochondrial surface area (C/A ratio), directly related to mitochondrial function. No statistically significant correlation (Pearson's test) was found between the two methods in any of the experiments. Only in a few micrographs were the values similar (n = 3) or very close (n = 2) to be comprised within the s.e.m. of their experimental group. Moreover, as judged by the s.d. comparison, the scatter of values was more prominent with the ML-operated than with the manual method. Conceivably, this outcome is because many ultrastructural details of the cell organelles are similar, for example, the membrane section profiles, and can only be properly recognized and distinguished by an experienced observer, while the current ML protocols still cannot.