Glioma Identification Based on Digital Multimodal Spectra Integrated with Deep Learning Feature Fusion Using a...

Area of Science:

• Biomedical Engineering
• Spectroscopy
• Artificial Intelligence

Background:

• Miniature fiber Raman spectroscopy offers non-invasive, real-time, label-free biomolecular analysis with potential for early cancer diagnosis.
• Glioma detection faces challenges due to spectral heterogeneity and low signal-to-noise ratios from miniature spectrometers, impacting model accuracy.
• Existing multimodal tumor detection methods often require multiple instruments, increasing hardware costs.

Purpose of the Study:

• To develop an accurate and cost-effective glioma identification method using miniature Raman spectroscopy.
• To enhance brain glioma recognition by integrating digital multimodal spectra with deep learning features fusion (DMS-DLFF).
• To improve intraoperative tumor detection instruments for portable and low-cost applications.

Main Methods:

• Mathematically decomposed original spectra into Raman and fluorescence components to augment biospectral information.
• Applied deep learning to extract features from both spectra, achieving feature-level digital multimodal spectral fusion.
• Constructed a two-layer pattern recognition model using an ensemble strategy with a bagging strategy to enhance Support Vector Machine algorithms.

Main Results:

• The proposed DMS-DLFF method achieved high accuracy (91.9%), sensitivity (96.7%), and specificity (80.8%) in brain glioma detection.
• Demonstrated superior performance compared to traditional algorithms in distinguishing glioma from normal brain tissue.
• Collected 260 Raman spectra of glioma and 151 of normal brain tissue for model training and validation.

Conclusions:

• The DMS-DLFF method effectively enhances brain glioma detection accuracy without additional hardware costs.
• This approach provides a promising tool for precise surgical planning and improved patient prognosis in neuro-oncology.
• Digital multimodal spectra integrated with deep learning offer a robust strategy for overcoming limitations in miniature spectroscopic analysis.