Comparison and optimization of machine learning methods for automated classification of circulating tumor cells

Area of Science:

• Biotechnology
• Medical Diagnostics
• Computational Biology

Background:

• Rare cell capture technology enables circulating tumor cell (CTC) analysis from patient blood.
• Manual cell counting is a bottleneck, being time-consuming, inconsistent, and prone to bias.

Purpose of the Study:

• To automate the location and classification of captured cells using machine learning (ML).
• To explore the impact of different ML algorithms on rare cell detection performance.

Main Methods:

• Trained four ML algorithms on three distinct datasets.
• Evaluated algorithm performance based on speed and accuracy (Area Under the ROC Curve).

Main Results:

• ML algorithms processed thousands of cells in minutes, a significant speed increase over manual methods.
• Some ML algorithms demonstrated superior performance (P < 0.05) compared to others.
• Training and testing discrepancies between cell lines and CTCs indicate a need for representative training data.

Conclusions:

• Automated ML approaches dramatically enhance rare cell detection efficiency and consistency.
• Algorithm selection and optimization are crucial and dataset-dependent for accurate rare cell classification.
• Future work should focus on training ML models with data representative of real-world clinical samples.