Benchmarking Machine Learning and Automated Image Analysis for Organelle Quantification

Abstract

Quantitative organelle analysis is highly sensitive to image-processing choices, limiting reproducibility across microscopy studies. Here, we systematically compare automated, interactive machine learning, and deep learning-based pipelines for lipid droplet and mitochondrial quantification in live human osteosarcoma cells imaged by fluorescence microscopy and label-free holotomography. Using standardized downstream feature extraction, we evaluated script-based workflows (Fiji, Python), a modular platform (CellProfiler), interactive machine learning (ilastik), and pretrained deep learning models. Lipid droplet segmentation was qualitatively consistent across approaches; however, droplet counts, and size distributions varied substantially between pipelines and imaging modalities, with ilastik reducing background-driven detections and improving cross-modality agreement. In contrast, mitochondrial quantification proved highly sensitive to segmentation and skeletonization choices, particularly in holotomography where global intensity-threshold-based methods failed to capture network structure. Based on these cross-pipeline comparisons, we demonstrate how organelle- and modality-specific benchmarking can guide pipeline selection, illustrated by the analysis of metabolic perturbations affecting lipid droplets and mitochondria. Together, these results highlight modality- and morphology-dependent limitations in common analysis pipelines and provide practical guidance for selecting robust, reproducible strategies for quantitative organelle imaging.