Deep learning classification of lipid droplets in quantitative phase images

Area of Science:

• Biomedical imaging
• Cell biology
• Machine learning

Background:

• Accurate lipid droplet classification is crucial for understanding cellular metabolism.
• Traditional imaging methods like fluorescence and Raman spectroscopy have limitations, including phototoxicity and the need for labels.

Purpose of the Study:

• To apply supervised machine learning for automated lipid droplet classification in label-free, quantitative-phase images.
• To compare the performance of various machine learning algorithms for this task.

Main Methods:

• Quantitative-phase imaging was used to acquire label-free images of single living cells.
• Several supervised machine learning algorithms were implemented and compared, including convolutional neural networks.
• Performance was evaluated based on accuracy, computational efficiency, and training resource requirements.

Main Results:

• Convolutional neural networks demonstrated superior performance in lipid droplet classification compared to other methods, both quantitatively and qualitatively.
• The developed approach enables accurate classification of lipid droplets in single living cells.
• The method offers an alternative to fluorescent and Raman imaging with ultra-low phototoxicity.

Conclusions:

• Quantitative-phase imaging combined with machine learning provides an effective and label-free method for lipid droplet classification.
• This paradigm offers significant advantages, including reduced phototoxicity and deeper insights into cellular metabolism thermodynamics.