A research on applying the diffusion model algorithm for Infrared and Raman spectroscopy data augmentation to improve...

In recent years, Raman and Infrared spectroscopy have become important tools in disease diagnosis due to their high sensitivity and non-invasive detection advantages. Combined with deep learning methods, spectral data can achieve high-precision identification of a wide range of diseases. However, deep learning models rely on large amounts of high-quality data. In practice, spectral data often faces challenges such as limited sample size, noise interference, and device variability, leading to model overfitting and poor generalization. To overcome this bottleneck, this paper proposes a spectral data generation method based on a diffusion model. By encoding temporal and category label information separately and combining it with a multi-head attention mechanism, the method extracts the overall morphology and local subtle features of the spectrum at multiple scales. In the reverse reconstruction stage, the noise distribution is estimated based on implicit modeling and cross-attention is used to generate spectra of different categories under different labels. This method achieves accurate denoising under conditional constraints and progressively reconstructs information such as the positions of characteristic spectral peaks. This method generates high-quality, continuous spectral data, effectively enhancing the scale and quality of the dataset. Experimental results demonstrate that the proposed method can generate high-fidelity and diverse synthetic spectral samples. The Pearson correlation coefficients for the thyroid and SLE infrared datasets reached high values of 0.97719 and 0.9914, respectively, and the similarities for the Thyroid and SLE Raman datasets reached high values of 0.9531 and 0.9747, respectively. This method effectively expands the training dataset and alleviates data scarcity and uneven distribution. The diffusion model was applied to two disease diagnosis tasks for validation. Under the condition of optimal gain ratio,the model trained with augmented data significantly improved generalization performance and diagnostic accuracy. In the infrared spectroscopy classification task, the accuracy of SLE disease on EfficientNet improved from 69.70% to 90.91%, and the accuracy of benign and malignant thyroid tumors on EfficientNet improved from 66.67% to 89.47%. In the Raman spectroscopy classification task, the accuracy of SLE disease on MLP improved from 87.88% to 90.91%, and the accuracy of benign and malignant thyroid tumors on the Transformer model improved from 87.22% to 92.98%.This study provides an effective generative augmentation framework for small-sample spectral data analysis, with strong theoretical value and application prospects.