Unsupervised convolutional variational autoencoder deep embedding clustering for Raman spectra.

Area of Science:

• Spectroscopy and Chemometrics
• Machine Learning and Data Mining
• Deep Learning for Scientific Applications

Background:

• Unsupervised deep learning methods are crucial for analyzing unlabeled data, with clustering algorithms widely used in various fields.
• Deep embedding clustering methods, while successful in image and speech recognition, have seen limited application in spectral data analysis.
• Raman spectra analysis often requires robust clustering techniques to differentiate samples based on subtle spectral features.

Purpose of the Study:

• To develop and evaluate an unsupervised deep learning clustering algorithm specifically designed for Raman spectra.
• To improve upon existing variational autoencoder-based clustering methods by adapting network architectures for spectral data.
• To demonstrate the efficacy of the proposed method in handling spectral datasets with minimal feature differences and small data sizes.

Main Methods:

• Introduced the convolutional variational autoencoder deep embedding clustering method (CVDE), an unsupervised algorithm for Raman spectra.
• Replaced the traditional multi-layer perception's fully connected layers with convolutional and pooling layers to enhance feature extraction for spectra.
• Integrated gradient-weighted class activation mapping (Grad-Cam) for visualizing spectral feature importance in clustering.

Main Results:

• CVDE achieved superior clustering performance compared to advanced methods on MNIST, soybean oil Raman spectra, and drug Raman spectra datasets.
• Achieved clustering accuracies of 94.48% (MNIST), 90.43% (soybean oil spectra), and 98.70% (drug spectra).
• Demonstrated effectiveness in handling datasets with very small Raman feature differences and limited data points.

Conclusions:

• CVDE is a versatile and effective unsupervised clustering method suitable for static spectra analysis, including Raman and LIBS spectra.
• The convolutional network structure improves feature learning and prevents overfitting in spectral clustering tasks.
• CVDE offers a promising alternative to supervised methods in spectral and chemical analysis, particularly for unlabeled data.