Artificial intelligence-enabled lipid droplets quantification: Comparative analysis of NIS-elements Segment.ai and ZeroCostDL4Mic StarDist networks.
Journal:
Methods (San Diego, Calif.)
PMID:
40023351
Abstract
Lipid droplets (LDs) are dynamic organelles that are present in almost all cell types, with a particularly high prevalence in adipocytes. The phenotype of LDs in these cells reflects their maturity, metabolic activity and function. Although LDs quantification in adipocytes is significant for understanding the origins of obesity and associated complications, it remains challenging and requires the implementation of computer science innovations. This article outlines a practical workflow for application of Segment.ai neural network from the commercial software NIS-Elements and the open-source StarDist Jupyter notebook from the ZeroCostDL4Mic platform for the analysis of LDs number and morphology. To generate a training dataset, 3T3-L1 cells were differentiated into adipocytes and stained with lipophilic dye BODIPY493/503. Subsequently, confocal live cell images were acquired, annotated and used for training. As an example task, deep learning models were tested on their ability to detect LDs enlargement on images of adipocytes with inhibited lipolysis. We demonstrated that both Segment.ai and StarDist models are capable of accurately recognising LDs on microphotographs, thereby significantly accelerating the processing of imaging data. The advantage of the Segment.ai model is its integration into NIS-Elements General Analysis 3, which performs quantitative and statistical data interpretation. Alternatively, StarDist is a more accessible and transparent tool, enabling precise model evaluation. In conclusion, both created approaches have the potential to accelerate the exploration of LDs dynamics, thus paving the way for further insights into how these organelles regulate energy homeostasis and contribute to the development of metabolic abnormalities.