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

Raman spectroscopy in biological applications faces challenges due to complex spectra, characterized by peaks of varying widths and significant biological background noise. Convolutional neural networks (CNNs) are widely used for spectrum classification due to their ability to capture local peak features. In this study, we introduce a multiscale CNN designed to detect weak biomolecule signals and differentiate spectra with features that cannot be statistically distinguished. The approach is further enhanced by a new visualization technique tailored for multiscale spectral analysis, providing clear insights into classification results. Using the classification of cholera toxin B subunit (CTB)-treated versus untreated cell membrane samples, whose spectra cannot be statistically differentiated, the optimized multiscale CNN achieved superior performance compared to traditional machine learning methods and existing multiscale CNNs, with accuracy (99.22%), sensitivity (99.27%), specificity (99.16%), and precision (99.20%). Our new visualization method, based on gradients of activation maps with respect to class scores, generates saliency scores that capture sample variations, with decision-making relying on consistently identified peak features. By visualizing the effects of different kernel sizes, Grad-AM highlights features at varying scales, aligning closely with spectral features and enhancing CNN interpretability in complex biomolecular analysis. These advancements demonstrate the potential of our method to improve spectral analysis and reveal previously hidden peaks in complex biological environments.