Raman spectroscopy and topological machine learning for cancer grading

Area of Science:

• Biomedical Engineering
• Computational Biology
• Spectroscopy

Background:

• Raman spectroscopy offers biochemical mapping for tumor tissue classification.
• Tumor grading requires accurate tissue differentiation based on molecular composition.

Purpose of the Study:

• To evaluate the efficacy of combining persistent homology and machine learning for Raman spectra classification in tumor grading.
• To develop an automated pipeline for selecting optimal topological features and machine learning classifiers for chondrosarcoma grading.

Main Methods:

• Extraction of topological features from Raman spectra using persistent homology.
• Training machine learning classifiers (specifically, a support vector classifier) with Betti Curve representations.
• Utilizing cross and leave-one-patient-out cross-validation for accuracy assessment.

Main Results:

• The combined approach achieved 81% validation accuracy and 90% test accuracy for binary classification of chondrosarcoma.
• High accuracy was maintained even with data acquired at different times and using different equipment.
• The Betti Curve representation combined with a support vector classifier demonstrated superior performance compared to existing literature.

Conclusions:

• Persistent homology and machine learning integration provides a robust method for Raman spectra-based tumor grading.
• The developed model shows potential for seamless integration into clinical settings for improved chondrosarcoma diagnosis.
• This technique offers a promising advancement in objective and automated tumor classification systems.