Generative AI Augments Raman Spectra for Glioma Classification

Machine learning applications in Raman spectroscopy-based diagnostics, particularly for glioma analysis, are often hampered by the scarcity and variability of biomedical datasets. This study investigates how deep generative augmentation can mitigate these limitations in small-cohort settings. Researchers analyzed Raman spectra from 58 glioma tumor samples, focusing on both binary IDH-status and 6-class methylation subtype classification.To tackle the dataset's limited size and imbalance, a conditional variational autoencoder ($\beta$-CVAE) was developed to generate class-conditioned synthetic Raman spectra. The generated data was evaluated in scenarios where models were trained solely on synthetic data, or on a combination of synthetic and real data, then tested on real data, all under a strict patient-isolated cross-validation protocol.While models trained exclusively on synthetic data underperformed those trained on real data, indicating a domain gap, augmenting real training data with synthetic spectra consistently improved classification performance across multiple models. This suggests that even with limited patient samples, generative models can capture enough underlying structure to provide valuable regularization for downstream classifiers, enhancing robustness in data-scarce biomedical applications.