Machine learning utilising spectral derivative data improves cellular health classification through hyperspectral...

Area of Science:

• Biophotonics and Spectroscopy
• Computational Biology and Machine Learning
• Cellular Metabolism and Pathology

Background:

• Objective differentiation of cellular metabolism is crucial for clinical applications such as tumor boundary definition and wound debridement.
• Existing in vitro spectral biomarkers for live vs. necrotic/apoptotic cells face challenges due to the complexity of biological processes.
• Sophisticated, objective classification methods are needed to accurately delineate different cellular states using spectroscopy.

Purpose of the Study:

• To evaluate the efficacy of machine learning clustering on various pre-processed hyperspectral imaging data for assessing cellular health.
• To compare the diagnostic utility of different data pre-processing techniques (raw, smoothed, background subtracted, spectral derivative) for cellular state differentiation.
• To identify optimal spectral data analysis methods for objective classification of healthy and traumatized cells.

Main Methods:

• Hyperspectral imaging (2500-3500 nm) was performed on healthy and traumatized cell samples using a portable prototype device.
• Machine learning clustering algorithms were applied to raw, smoothed, background-subtracted, and spectral derivative processed data.
• Diagnostic performance was quantified using sensitivity and specificity to compare analysis methods.

Main Results:

• Raw spectral data showed limited differentiation among trauma types, with signal contamination hindering accurate clustering.
• Smoothed and background-subtracted data sets reduced accuracy by removing key spectral features indicative of cellular health.
• Spectral derivative data significantly improved clustering accuracy, achieving >94% sensitivity and specificity for background-subtracted data, effectively handling signal contamination.

Conclusions:

• Spectral derivative processing of hyperspectral imaging data is a highly effective method for objective cellular health assessment.
• This approach enhances machine learning-based classification of cellular states, overcoming challenges posed by signal contamination.
• The findings highlight the utility of spectral derivatives for accurate differentiation in clinical applications requiring precise cellular analysis.