PI-AttnGAN: Physics-informed attention generative modeling for cross-modal FT-IR-to-Raman serum spectral reconstruction.

Fourier-transform infrared (FT-IR) and Raman spectroscopy provide complementary molecular vibrational information and therefore hold considerable promise in rapid cancer diagnosis. However, Raman spectroscopy signals are weak and susceptible to noise interference, and the equipment cost is higher than that of FT-IR, which hinders the widespread clinical application of multimodal spectroscopic screening. Therefore, this paper proposes an FT-IR-to-Raman cross-modal generation model PI-AttnGAN that integrates physical constraints and self-attention mechanism, which is used to generate corresponding Raman spectra from readily acquired FT-IR. This model uses a one-dimensional U-shaped generator to extract multi-scale spectral features, introduces self-attention mechanism at the bottleneck layer to enhance long-range dependency modeling ability, and constrains both the overall spectral shape and local peak structure through a dual-scale conditional discriminator. In addition, the model introduces gradient difference loss and spectral angle mapper (SAM) loss to preserve local peak-shape variations and overall spectral vector direction of Raman spectra. The model was validated on paired serum infrared and Raman spectra from five sample groups, including healthy control group, glioma, renal cell carcinoma, lung cancer, and esophageal cancer. The results showed that the Raman spectra generated by PI-AttnGAN were able to maintain the main peak positions, global spectral profiles, and spectral distribution characteristics of the real Raman spectra well. The average root mean square error of the five types of samples was 0.0325, Pearson correlation coefficient was 0.9758, signal-to-noise ratio was 16.03 dB, spectral angle mapper was 0.1787 rad, Kullback-Leibler divergence was 0.0610, and the overall performance was better than the comparative models such as VAE, CNN-Bridge, and DDPM. Downstream classification experiments further demonstrated that incorporating the generated Raman spectra improved diagnostic performance, indicating that the generated spectra retained cancer-related discriminative information. These findings suggest that PI-AttnGAN provides an effective computational strategy for supplementing unavailable Raman measurements and may facilitate low-cost multimodal spectroscopic cancer screening.