High-Throughput Recognition of Tumor Cells Using Label-Free Elemental Characteristics Based on Interpretable Deep...

Area of Science:

• Biomedical Spectroscopy
• Cancer Diagnostics
• Machine Learning in Medicine

Background:

• Cancer poses a significant threat to increasing life expectancy.
• Instantaneous diagnostic methods are crucial for effective cancer management.
• Current diagnostic methods may lack the speed and throughput required for early detection.

Purpose of the Study:

• To develop a label-free laser-induced breakdown spectroscopy (LIBS) method for high-throughput tumor cell recognition.
• To establish a deep learning model for simultaneous classification of various cancer types.
• To enhance the interpretability of the deep learning classification using gradient-weighted class activation mapping.

Main Methods:

• Laser-induced breakdown spectroscopy (LIBS) was employed for label-free analysis of cancer cells.
• Cells were deposited on a silicon substrate for spectral data acquisition.
• A deep learning model, specifically a one-dimensional convolution neural network (1D-CNN), was developed for classification.
• Gradient-weighted class activation mapping (Grad-CAM) was used to interpret the 1D-CNN model.

Main Results:

• The 1D-CNN model achieved high performance metrics for cancer cell classification.
• Mean sensitivity reached 94.00%, mean specificity was 98.47%, and mean accuracy was 97.56%.
• Saliency maps generated by Grad-CAM effectively highlighted spectral differences between cancer cell lines, aiding interpretability.

Conclusions:

• The proposed label-free LIBS method offers a satisfactory and interpretable approach for cancer cell line classification.
• This technique demonstrates potential for high-throughput screening and rapid cancer diagnostics.
• The integration of LIBS with interpretable deep learning provides a powerful tool for biomedical research and clinical applications.