Supervised machine learning versus expert assessment of ultrastructural changes in wild-type and OGT knockout macrophages

Abstract

Automated transmission electron microscopy (TEM) generates large datasets that challenge traditional qualitative analysis of cellular ultrastructure. Quantitative assessment of structural differences between different samples remains difficult due to structural variability in thin sections of organelles. Here we applied supervised machine learning (sML) to segment, quantify and compare cellular structures -including nuclei, chromatin, mitochondria, rough endoplasmic reticulum, and endocytic vesicles- in large TEM images of wild-type macrophages versus those with altered cellular physiology due to deficiency in O-GlcNAc Transferase (OGT). sML revealed that OGT knockout macrophages are larger and more oval, with increased euchromatin, nucleoli size, and relative mitochondrial and rER surface areas. Comparison with six TEM experts showed sML provides more objective and sensitive quantification of subtle differences, while expert consensus is only achieved for larger structural variations. These findings demonstrate that sML enhances quantitative TEM analysis and complements human expertise in ultrastructural studies.