SCDM-Net: Symbiotic causal debiasing for SERS-based exosome profiling in lung cancer

Early detection of lung cancer improves patient prognosis. Surface-enhanced Raman scattering (SERS) analysis of exosomes provides an approach for lung cancer-related molecular profiling; however, instrument- and batch-related variation can obscure class-relevant spectral features. We developed a Symbiotic Causal Debiasing Multi-scale Network (SCDM-Net) to address this problem. A "symbiotic-parasitic" module uses gated residual fusion to model interactions between bias-associated and class-relevant representations. A dual-stream architecture processes raw one-dimensional spectra and two-dimensional short-time Fourier transform (STFT) representations using multi-scale convolutions to capture complementary spectral patterns across the fingerprint region and regions associated with lipids and proteins. A gradient reversal layer (GRL) encourages the separation of class-relevant and batch-associated representations. Using 4006 exosome SERS spectra from 12 independent biological batches, SCDM-Net achieved an accuracy of 90.77% and an F1 score of 0.8980 on a held-out test set comprising four biological batches not used during training. These proof-of-concept results, obtained using cell-line-derived exosomes, provide a methodological framework for SERS-based exosome profiling in lung cancer; clinical translation will require validation using patient-derived samples.