DeepLuAd: Semantic-guided virtual histopathology of lung adenocarcinoma via stimulated Raman scattering

Rationale: Accurate histologic grading of lung adenocarcinoma is essential for guiding clinical management. Conventional hematoxylin and eosin (H&E) staining provides morphological information but lacks biochemical specificity, limiting quantitative analysis of tissue subtypes within the heterogeneous lung cancer microenvironments. Methods: We developed DeepLuAd, an AI-powered platform integrating label-free stimulated Raman scattering (SRS) microscopy with semantic-guided deep learning. The platform enables automated tumor grading, segmentation, cellular-level morpho-chemical quantification, and unsupervised virtual H&E staining. Results: DeepLuAd achieved a mean intersection-over-union (mIoU) of 80.43% across major lung tissue subtypes and reached a grading concordance rate of 76.2% with pathologist diagnoses (16/21 cases). The approach also enabled quantitative mapping of lipid-to-protein ratio heterogeneity within tumor and stromal compartments, revealing biochemical signatures of disease progression. Conclusions: DeepLuAd provides an interpretable and scalable framework for digital lung adenocarcinoma analysis, unifying morphological and biochemical information without the need for staining. The method demonstrates potential for broader application to other solid tumors in AI-enhanced histopathology.