Data augmentation method based on the Gaussian kernel density for glioma diagnosis with Raman spectroscopy

Area of Science:

• Biomedical Spectroscopy
• Computational Pathology
• Machine Learning in Oncology

Background:

• Glioma, a highly infiltrative brain tumor, presents challenges in precise boundary identification during surgery.
• Raman spectroscopy offers potential for *in situ* and *in vivo* glioma boundary detection.
• Class imbalance due to limited normal tissue samples hinders the development of robust classification models.

Purpose of the Study:

• To propose a data augmentation algorithm, GKIM, for enhancing normal brain tissue spectra in Raman spectroscopy datasets.
• To improve the robustness and accuracy of classification models for glioma detection by addressing class imbalance.

Main Methods:

• Developed a Gaussian kernel density-based algorithm (GKIM) for data augmentation of normal brain tissue Raman spectra.
• Introduced a Gaussian density-based weight coefficient calculation for synthesizing diverse new spectra.
• Employed fuzzy nearest neighbor distance for adaptive selection of spectra and synthesis, avoiding concentrated sample distribution.

Main Results:

• Augmented 136 normal brain tissue spectra to 600, effectively balancing the dataset with 769 glioma spectra.
• Achieved high classification performance with 91.67% accuracy, sensitivity, and specificity.
• Demonstrated superior predictive performance compared to traditional algorithms for imbalanced class data.

Conclusions:

• The GKIM algorithm effectively increases sample diversity and model robustness for brain tumor classification using Raman spectroscopy.
• Adaptive spectral synthesis using fuzzy nearest neighbor distance overcomes limitations of common data augmentation methods.
• This approach significantly improves diagnostic performance in cases of imbalanced spectral data for glioma detection.