Learning deep features for dead and living breast cancer cell classification without staining

Area of Science:

• Computational biology
• Cancer research
• Machine learning applications

Background:

• Automated cell classification is crucial in cancer biology and computer vision.
• Breast cancer exhibits diverse cell populations and stroma, complicating analysis.
• Automated microscopy enables large-scale live-cell imaging for database compilation.

Purpose of the Study:

• To classify breast cancer cells as live or dead using image processing and machine learning.
• To test the hypothesis that live-dead classification is possible without staining, using only bright-field images.
• To evaluate the performance of convolutional neural networks (CNNs) for this classification task.

Main Methods:

• Utilized the JIMT-1 breast cancer cell line.
• Compiled a large image dataset of cells treated with chemotherapy or vehicle control.
• Trained CNN classifiers using bright-field images and fluorescence microscopy labels.
• Evaluated classifier performance on a large set of bright-field images.

Main Results:

• The best model achieved an AUC of 0.941 for classifying untreated breast cancer cells.
• The model reached an AUC of 0.978 for classifying drug-treated breast cancer cells.
• Analysis linked classifier clusters to observable visual characteristics in live-dead cell biology.

Conclusions:

• Machine learning and computational image analysis show potential for developing new diagnostic tools.
• These methods can reduce costs, save time, and improve the reproducibility of biomedical research.
• The study successfully demonstrated label-free live-dead cell classification using bright-field microscopy and CNNs.