Predictive data clustering of laser-induced breakdown spectroscopy for brain tumor analysis

Area of Science:

• Spectroscopy
• Data Science
• Biomedical Engineering

Background:

• Supervised learning methods for laser-induced breakdown spectroscopy (LIBS) face limitations in diagnostic accuracy due to insufficient training spectral data across diverse tissues.
• Accurate tissue identification using LIBS is crucial for various diagnostic applications but is hindered by data scarcity.

Purpose of the Study:

• To improve the diagnostic accuracy of LIBS for tissue identification by integrating predictive data clustering with supervised learning.
• To evaluate the effectiveness of the meanshift clustering method against other clustering techniques in the LIBS field.
• To introduce a new evaluation metric, the cluster precision (CP) score, alongside the Calinski-Harabasz (CH) score for assessing clustering performance.

Main Methods:

• Implemented and compared the meanshift clustering method with three other established clustering methods for LIBS data.
• Introduced and utilized the cluster precision (CP) score in conjunction with the Calinski-Harabasz (CH) score for comprehensive cluster evaluation.
• Analyzed the impact of principal component analysis (PCA) on the performance of four different clustering algorithms.

Main Results:

• The PCA-Meanshift clustering method demonstrated superior performance based on a combined evaluation of CH and CP scores.
• PCA-Meanshift effectively leveraged spatial location and feature similarity information to enhance predictive clustering.
• Integration of PCA-Meanshift improved diagnostic accuracy from below 95% to 100% across multiple classifiers, including SVM, k-NN, Simca, and RF.

Conclusions:

• The proposed PCA-Meanshift clustering approach significantly overcomes the limitations of traditional supervised learning in LIBS tissue identification.
• This hybrid method provides a robust framework for achieving perfect diagnostic accuracy in LIBS applications.
• The developed CP score serves as a valuable metric for evaluating clustering effectiveness in spectroscopic data analysis.