Interpretable Multiscale Convolutional Neural Network for Classification and Feature Visualization of Weak Raman...

Area of Science:

• Biomolecular analysis
• Spectroscopy
• Machine learning

Background:

• Raman spectroscopy in biology presents challenges due to complex spectra and background noise.
• Convolutional neural networks (CNNs) are effective for spectrum classification by capturing local peak features.

Purpose of the Study:

• To develop a multiscale CNN for detecting weak biomolecule signals in complex spectra.
• To enhance spectral differentiation capabilities beyond statistical distinguishability.
• To introduce a novel visualization technique for improved interpretability of multiscale spectral analysis.

Main Methods:

• Implementation of a multiscale CNN architecture for spectral feature extraction.
• Application of a new gradient-based activation map visualization technique (Grad-AM) for interpretability.
• Validation using cholera toxin B subunit (CTB)-treated versus untreated cell membrane samples.

Main Results:

• The optimized multiscale CNN achieved high performance: 99.22% accuracy, 99.27% sensitivity, 99.16% specificity, and 99.20% precision.
• The method successfully differentiated spectra that were statistically indistinguishable.
• The Grad-AM visualization highlighted key spectral features at various scales, aligning with CNN decision-making.

Conclusions:

• The developed multiscale CNN and visualization technique offer superior performance in complex biological spectral analysis.
• This approach enhances the detection of weak biomolecular signals and improves the interpretability of CNN models.
• The method holds significant potential for advancing spectral analysis in challenging biological environments.