Microfluidic droplets content classification and analysis through convolutional neural networks in a liquid biopsy...

Area of Science:

• Oncology
• Biotechnology
• Computational Biology

Background:

• Liquid biopsy enables detection of circulating tumor cells (CTCs) in peripheral blood using microfluidics and extracellular acidification (ECAR) measurements.
• Current methods rely on manual droplet screening, increasing operator dependency and assay time.
• Variations in droplet volume can lead to inaccurate ECAR estimations.

Purpose of the Study:

• To integrate computer vision and AI for automated analysis of microfluidic liquid biopsy.
• To improve the accuracy of ECAR measurements by accounting for droplet and cell volume variations.
• To explore additional information within droplet images for enhanced cancer detection.

Main Methods:

• Convolutional neural networks (CNNs) were implemented for automatic classification of droplets, achieving over 96% accuracy.
• Object detection neural networks were used to segment droplets and cells, enabling precise volume measurement and ECAR correction (up to 20%).
• Machine learning models analyzed bright-field cell images to predict CD45 expression on white blood cells (82.9% accuracy) and classify acid droplets from breast cancer patients (90.2% accuracy).

Main Results:

• Automated droplet classification significantly reduced operator dependency and assay time.
• Accurate ECAR measurements were achieved by correcting for volume variations.
• The AI-driven approach demonstrated high accuracy in classifying metastatic breast cancer patient samples.

Conclusions:

• Computer vision and AI integration substantially improve the efficiency and accuracy of microfluidic liquid biopsy for CTC detection.
• The refined ECAR measurement and image analysis offer a more robust method for cancer diagnostics.
• This AI-enhanced liquid biopsy shows promise for accurate classification of metastatic breast cancer patients.