Tumor grading drives treatment decisions. Accurate grading of lung adenocarcinoma - which accounts for nearly half of all lung cancer cases globally - is essential for deciding how aggressively to treat each patient. Current grading guidelines from the International Association for the Study of Lung Cancer (IASLC) require pathologists to estimate what percentage of the tumor shows high-grade patterns, with a 20% threshold determining whether a tumor is classified as high-grade (Grade 3).
H&E staining has inherent limitations. The gold standard for tissue analysis, hematoxylin and eosin (H&E) staining, provides only morphological information about cell shapes and tissue architecture. It reveals nothing about the biochemical composition of cells, and the grading process relies heavily on a pathologist's subjective visual judgment. This subjectivity leads to variability between pathologists and limits reproducibility.
Heterogeneity makes grading difficult. Lung adenocarcinoma tumors frequently contain mixtures of different histological subtypes - from low-grade lepidic patterns to high-grade solid and micropapillary patterns - all within the same tissue specimen. Precisely quantifying the proportion of each subtype requires painstaking manual analysis that is slow, tedious, and error-prone at scale.
An opportunity for label-free imaging combined with AI. Stimulated Raman scattering (SRS) microscopy offers a fundamentally different approach: it can image unstained tissue by detecting molecular vibrations specific to lipids, proteins, and collagen without any chemical staining. When combined with artificial intelligence, SRS could provide faster, more objective, and biochemically richer tissue analysis than conventional methods.
A three-module AI framework. DeepLuAd is an integrated artificial intelligence platform that combines label-free SRS microscopy with deep learning to perform three tasks simultaneously: tissue segmentation (SegLuAd), biochemical and cellular quantification (QuantLuAd), and virtual H&E staining (VStainLuAd). Together, these modules transform raw optical images into clinically interpretable histopathology without any staining.
SRS imaging captures molecular contrast. The SRS microscopy system detects vibrational signals at two frequencies corresponding to lipid (CH2 bonds at 2845 cm-1) and protein (CH3 bonds at 2930 cm-1). A third channel captures second harmonic generation (SHG) signals from collagen fibers. The resulting color-coded composite image - red for collagen, green for lipid, blue for protein - provides the raw data for all downstream AI analysis.
From chemistry to diagnosis. Rather than just classifying whether tissue is cancer or not, DeepLuAd performs fine-grained, pixel-level segmentation of seven distinct tissue classes including three grades of adenocarcinoma. This granularity is what allows the platform to apply the same quantitative grading rules that pathologists use, making the AI decision-making process transparent and clinically meaningful.
Designed for scalability and extension. The modular architecture of DeepLuAd is designed to be adaptable to other solid tumor types beyond lung adenocarcinoma. The same framework - SRS imaging combined with semantic segmentation and quantification - could in principle be applied to breast, pancreatic, or prostate cancer, offering a versatile platform for next-generation digital pathology.
Seven tissue types captured by SRS. SRS microscopy successfully distinguished all major lung tissue types - including normal alveoli, immune cell aggregates, and six histological subtypes of lung adenocarcinoma ranging from low-grade lepidic patterns to high-grade solid and complex glandular patterns. The morphological features visible in SRS images closely matched those seen in paired conventional H&E-stained sections from the same tissue.
Biochemical contrast aids subtype identification. Beyond morphology, SRS provides intrinsic biochemical information that further distinguishes tissue subtypes. Normal alveoli show low lipid content, while tumor cells display progressively elevated lipid and protein signals. Collagen-rich stromal regions produce strong SHG signals, helping demarcate tumor boundaries from surrounding supportive tissue.
High-grade cancer has distinctive chemical signatures. Even at the molecular level, higher-grade tumor subtypes like solid predominant adenocarcinoma show markedly different lipid-to-protein ratios and cell packing compared to low-grade lepidic adenocarcinoma. These chemical differences are not visible in conventional H&E staining, representing a genuine information advantage of SRS-based analysis.
Working with frozen rather than fresh tissue. Because lung tissue contains air-filled alveolar cavities that deform easily, the team used thin frozen tissue sections rather than fresh whole-tissue imaging. Adjacent sections were prepared - one for standard H&E staining used as a reference by pathologists, and one kept unstained for SRS imaging - allowing direct comparison of the two imaging approaches.
Transformer-based architecture for pixel-level precision. The segmentation module SegLuAd is built on the Swin Transformer backbone, a state-of-the-art architecture that excels at capturing both local and long-range spatial relationships in images. This is paired with a UPerNet decoder to produce precise, pixel-level segmentation maps across all seven tissue classes - a significant improvement over earlier patch-level classification approaches that lacked spatial resolution.
Training data and annotation process. Pathologists annotated 80 SRS slides from 50 patients, manually delineating tissue regions using both the SRS image and the adjacent H&E-stained section as a morphological reference. These pixel-level annotations required expert knowledge but ensured high-quality ground truth. The model was trained on 75% of slides and validated on 25%, with an independent test cohort of 21 cases used for final evaluation.
Strong segmentation performance. SegLuAd achieved a mean intersection-over-union (mIoU) score of 80.43% across all tissue classes - a metric that measures how well the predicted tissue regions match the expert annotations. Stroma and high-grade tumor regions showed the highest accuracy (IoU above 85%), while immune cells were the most challenging class (IoU 70.97%). The model outperformed established segmentation architectures including U-Net, DeepLabV3, and PSPNet.
Automatic tumor grading from segmentation maps. Once segmentation is complete, DeepLuAd applies IASLC grading rules automatically: if high-grade patterns exceed 20% of total tumor area, the case is classified as Grade 3; otherwise, the dominant subtype determines the grade. Applied to 21 independent cases, this pipeline achieved 76.2% concordance with pathologist diagnoses (16 of 21 cases matched), demonstrating real clinical utility.
Lipid content increases with tumor grade. QuantLuAd measured lipid signal intensity across all tissue types and grades. Normal alveoli had the lowest lipid content, while lipid intensity progressively increased from low-grade (109.8 arbitrary units) to intermediate-grade (127.2) to high-grade (138.3) adenocarcinoma. This likely reflects cancer-specific lipid metabolic reprogramming, where tumor cells accumulate lipids to fuel rapid growth and membrane production.
Protein content is elevated across all tumor types. Protein signal was similarly elevated in all tumor subtypes compared to normal alveoli. High protein content correlates with the dense cellular packing and high biosynthetic demand of cancer cells, which must continuously produce new proteins to support rapid division. Stroma, immune cells, and tracheal wall showed lower protein content than cancer regions.
Cell morphology changes with grade. Nuclear-to-cytoplasm ratios increased progressively from low-grade (0.18) to high-grade (0.21) tumors, reflecting the classic pathological hallmark of cancer where cell nuclei enlarge relative to the surrounding cytoplasm. Average cell size decreased from 221.5 square micrometers in low-grade to 198.0 in high-grade tumors, while cell density in alveolar spaces increased - both consistent with the more aggressive, poorly differentiated biology of high-grade cancer.
AI quantification matches pathologist annotations. When the researchers compared biochemical measurements derived from AI segmentation versus those from manual expert annotations, they found near-perfect agreement - Spearman correlation and concordance correlation coefficients of approximately 0.99. This confirms that SegLuAd's segmentation is accurate enough to support reliable downstream biochemical quantification without requiring hand-annotated regions.
Converting SRS images to familiar H&E appearance. Despite the advantages of SRS imaging, pathologists are trained to interpret the purple-pink appearance of conventional H&E stained sections. VStainLuAd addresses this barrier by using an AI generative model (CycleGAN) to translate SRS images into virtual H&E images that look like standard clinical histology, without requiring any paired SRS-H&E image pairs for training.
Semantic guidance improves virtual staining quality. Standard CycleGAN models tend to produce blurring and color distortions in complex tissue areas. VStainLuAd solves this by feeding the segmentation maps from SegLuAd directly into the image generation process as 'semantic priors.' Different tissue types receive different attention weights, helping the model accurately reconstruct the distinct staining characteristics of each region - for example, deeply stained nuclei in high-grade tumor cells versus light cytoplasmic staining in low-grade glandular structures.
Transfer learning from brain tissue SRS data. The virtual staining model was first pretrained on brain SRS images, which have more consistent molecular distributions, then fine-tuned on lung tissue data. This two-stage training strategy improved convergence speed and final image quality compared to training on lung data alone, demonstrating that knowledge from other tissue types can usefully transfer to lung cancer analysis.
Pathologist-validated quality approaching real H&E. In a blinded evaluation by two senior pathologists using a 10-point quality scale, VStainLuAd virtual staining achieved a mean score of 9.66, compared to 9.68 for real FFPE H&E sections and only 6.20 for conventional CycleGAN without semantic guidance. This near-perfect quality score confirms that DeepLuAd's virtual staining is clinically interpretable and approaches the gold standard.
Advancing beyond patch-level AI. Most previous AI systems for SRS-based histopathology classified small image patches as either tumor or normal, lacking the spatial detail needed to distinguish between histological subtypes within a single lesion. DeepLuAd's pixel-level segmentation-first approach allows it to capture the heterogeneous mixture of subtypes within individual tumors - which is precisely what clinicians need to grade cancer accurately.
Biochemical metrics as future diagnostic markers. The lipid-to-protein ratio and cell density metrics revealed by QuantLuAd show clear associations with tumor grade. If validated in larger cohorts, these quantitative biochemical measures could serve as objective diagnostic markers that complement or partially replace subjective visual assessment, potentially reducing inter-pathologist variability in grading decisions.
Current practical limitations. The study has several limitations that must be addressed before clinical adoption. The dataset came from a single medical center with a relatively small number of patients, which may limit generalization to diverse populations. SRS microscopy requires specialized, expensive equipment and trained operators. Manual pixel-level annotation is extremely labor-intensive, limiting rapid dataset expansion. Future work will explore semi-automated annotation and multi-center validation.
A path toward intraoperative diagnosis. SRS microscopy's key advantage is speed - it can image fresh, unprocessed tissue sections within minutes, unlike conventional H&E staining which requires hours of fixation and processing. DeepLuAd could therefore enable same-session intraoperative histopathology, helping surgeons make real-time decisions about resection margins during lung cancer surgery without waiting for post-operative pathology reports.
A unified platform for next-generation histopathology. DeepLuAd successfully demonstrates that label-free SRS microscopy combined with multi-task AI can perform all three essential functions of clinical histopathology - tissue classification, biochemical analysis, and image generation for pathologist review - without any staining reagents or tissue processing beyond cryosectioning.
Grading concordance with pathologists. The 76.2% grading concordance rate with expert pathologists in an independent cohort is clinically meaningful, especially considering that even experienced human pathologists show significant inter-observer variability in grading heterogeneous tumors. Future model improvements with larger datasets are expected to push this concordance substantially higher.
New insights into tumor biochemistry. Beyond diagnostic utility, DeepLuAd reveals quantitative biochemical differences between cancer grades that are invisible to conventional histology. The progressive increase in lipid content and cell density from low to high-grade tumors opens new research avenues exploring lipid metabolism as a driver of lung cancer aggressiveness and a potential therapeutic target.
Broader applicability in cancer diagnostics. The modular framework of DeepLuAd is designed to extend to other cancers where SRS histopathology has already shown promise, including brain, breast, prostate, and gastrointestinal tumors. This scalability positions label-free AI histopathology as a transformative technology for faster, more objective, and more information-rich cancer diagnosis across multiple oncology specialties.