Artificial intelligence in Raman-based tumor histopathology: From spectral preprocessing to computational morphochemical imaging

Abstract

Raman-based tumor histopathology is evolving from isolated spectral classification toward artificial intelligence (AI)-enabled morphochemical imaging that integrates tissue morphology, molecular composition, and spatial context. This review examines AI applications across spontaneous Raman microscopy, line-scanning and hyperspectral Raman imaging, coherent anti-Stokes Raman scattering, stimulated Raman scattering/stimulated Raman histology (SRS/SRH), and surface-enhanced Raman scattering (SERS). Rather than cataloguing instruments or algorithms, we establish a task-oriented framework linking measurement physics, spectral-spatial data structure, pathological label granularity, model architecture, validation unit, and intended clinical decision. We discuss how chemometrics, convolutional and spatial-spectral networks, multiple-instance learning, self-supervised and contrastive learning, foundation models, and generative virtual staining address tumor detection, grading, molecular subtyping, biopsy assessment, infiltration mapping, and surgical-margin evaluation. Current evidence indicates that SRS/SRH has achieved the greatest translational maturity in brain and central nervous system tumors, owing to its compatibility with fresh tissue, rapid acquisition, histology-like output, and emerging prospective validation. By contrast, spontaneous and hyperspectral Raman approaches offer richer biochemical attribution and large-area mapping but remain constrained by acquisition speed, preprocessing sensitivity, instrument dependence, and patient-limited datasets. SERS provides highly sensitive multiplexed molecular imaging, yet its reliability and regulatory translation are inseparable from nanoprobe chemistry, delivery, nonspecific binding, and batch standardization. Across modalities, major barriers include data leakage, spatial autocorrelation, domain shift, insufficiently matched clinical endpoints, limited uncertainty reporting, and explanations that are chemically or pathologically incomplete. We therefore propose that clinical-grade AI-Raman pathology requires standardized preprocessing, patient-or specimen-level partitioning, external and prospective workflow validation, calibrated uncertainty and abstention, systematic failure analysis, and pathologist-reviewed spectral-spatial interpretability.