The Heterogeneity Challenge. Lung cancer remains the deadliest cancer type, with an estimated 234,580 new cases and approximately 340 deaths per day in the United States in 2024. A key driver of treatment failure is tumor heterogeneity -- the presence of spatially and functionally distinct cell populations within the same tumor. These populations fuel drug resistance, metastatic spread, and disease progression. Understanding how these populations evolve over time is essential for improving outcomes.
The Single-Cell RNA Sequencing Limitation. The gold standard for studying cell differentiation trajectories is single-cell RNA sequencing (scRNA-seq), which orders cells along developmental pathways based on gene expression profiles. While powerful, scRNA-seq is expensive, requires specialized equipment, demands significant tissue material, and is inaccessible in resource-limited settings. These barriers prevent large-scale application to routine clinical samples.
A Cost-Effective Alternative. Hematoxylin and eosin (H&E) stained slides are produced routinely as part of standard cancer diagnosis and represent the most widely available pathology data in oncology. The researchers developed an AI framework that extracts cell morphology features and spatial organization from these standard slides to classify cell differentiation status, infer cell dynamic trajectories, and quantify tumor progression -- all without any additional molecular testing.
Study Scope. The framework was validated in three independent lung adenocarcinoma cohorts: the National Lung Screening Trial (NLST-ADC, 148 patients), TCGA-LUAD (167 patients), and the University of Texas SPORE in Lung Cancer (52 patients). Integration with spatial transcriptomic and bulk RNA sequencing data provided molecular validation of the image-derived metrics.
Five-Class Prediction Model. The framework begins with a deep learning model trained to classify individual image patches into five categories: G1 (well-differentiated), G2 (moderately differentiated), G3 (poorly differentiated), G4 (undifferentiated), and an 'other' class covering normal tissue, necrosis, and other pathological conditions. Tumor grades reflect how closely cells resemble normal lung tissue -- low-grade tumors (G1) maintain structured organization and grow slowly, while high-grade tumors (G3, G4) are disorganized and aggressive.
Foundation Model Fine-Tuning. The Phikon histopathology foundation model -- pre-trained specifically on H&E images from TCGA -- was selected as the backbone after comparing five pretrained models. Phikon outperformed Phikon-V2 and all ResNet variants (ResNet-50, 101, 152) on the cell differentiation task, achieving a weighted accuracy of 0.61 at the patch level (Cohen's kappa 0.468) compared to ResNet-152's 0.41. The advantage stems from Phikon's domain-specific pre-training on pathology images versus ResNet's training on natural photograph datasets.
Whole-Slide Image Validation. At the whole-slide image (WSI) level, the model achieved a weighted accuracy of 0.70 on 148 slides. The AUROC was 0.89 for distinguishing low-grade from intermediate/high-grade tumors and 0.83 for distinguishing low/intermediate-grade from high-grade tumors. Expert pathologists validated the model's predicted regions of interest, confirming alignment with annotated tumor regions. The primary classification challenges involved the G2/G3 boundary, which even experienced pathologists find difficult to distinguish.
Pseudotime Analysis from Image Features. Analogous to how scRNA-seq uses gene expression to order cells along trajectories, this framework uses image-derived morphological features to estimate pseudotime. The pretrained model scans the entire slide pixel by pixel, extracting feature vectors from each patch. These features are clustered using the Leiden algorithm, and pseudotime is estimated using the diffusion pseudotime algorithm. This process places image patches in a developmental timeline, revealing how tumor cells have evolved spatially across the slide.
The Fitness Score Framework. Three complementary metrics were derived from the AI model outputs. Progression speed (S) quantifies how rapidly tumor cell differentiation advances spatially, reflecting biological aggressiveness. Shannon diversity (I) measures morphological heterogeneity across tumor cell populations. The fitness score combines both components as their product, capturing both aggressiveness and clonal diversity simultaneously. These metrics can be calculated from any H&E slide without additional molecular testing.
Consistent Survival Stratification Across Three Cohorts. Patients divided by median fitness score into 'slow' and 'fast' progression groups showed statistically significant survival differences in all three independent datasets: NLST (p = 0.01), LUAD (p = 0.02), and SPORE (p = 0.03). Progression speed also significantly stratified patients in all three cohorts (NLST p = 0.04, LUAD p = 0.02, SPORE p = 0.04). Shannon diversity was significant in NLST (p = 0.05) and LUAD (p = 0.05). The slow progression group consistently showed better overall survival across all metrics and datasets.
Improving Clinical Prognostic Models. Adding image-derived scores to standard clinical factors (cancer stage, gender, age) in multivariate Cox regression improved prognostic accuracy. The concordance index increased from 0.67 to 0.70 for the NLST cohort and from 0.72 to 0.82 for the SPORE cohort after adding image features. The fitness score proved more robust than speed or Shannon diversity individually, showing minimal sensitivity to the number of principal components used -- an important property for clinical standardization.
Three-Group Stratification. Combining speed and Shannon diversity to classify patients as 'extremely slow' (below median for both), 'extremely fast' (above median for both), or 'moderate' further refined survival prediction beyond the simple two-group split. This three-tier stratification demonstrated clinically meaningful differences across all three datasets, with the extremely fast group showing markedly worse outcomes than both moderate and slow groups.
Fast Progression: Cell Cycle Overactivation. Gene set enrichment analysis (GSEA) of transcriptomic data from SPORE and LUAD cohorts revealed that fast-progressing tumors (high Shannon diversity, high speed) show upregulation of cell cycle pathways including DNA replication, synthesis phase, and G2/M checkpoint. This molecular profile corresponds to increased cell division and proliferation, confirming that the image-derived fast progression classification captures genuine biological aggressiveness rather than arbitrary image features.
Slow Progression: Retained Normal Lung Identity. In contrast, slow-progressing tumors showed higher expression of surfactant metabolism genes and were significantly enriched in alveolar type 2 (AT2) cell signatures. AT2 cells are responsible for surfactant production, alveolar repair, and innate immune regulation in the normal lung. Slow-progression tumors retaining AT2 cell characteristics are more differentiated, grow more slowly, and behave less aggressively. The AT2 score was negatively correlated with speed, diversity, and fitness scores in both the SPORE and LUAD datasets.
Key Genes Associated with Progression. Specific genes were identified as markers of fast versus slow progression. VEGFB was upregulated in tumor regions and drives metastasis, making it a potential therapeutic target. PRR13 correlated with poor survival in adenocarcinoma. LTA4H reflected inflammatory signaling and cell proliferation. For slow progression, surfactant proteins SFTPA, SFTPB, and SFTPC were characteristic, along with CXCL17 and CD81 -- the latter expressed on B cells and involved in anti-tumor activity. MAFF acted as a tumor suppressor by inhibiting cell cycle progression.
Spatial Transcriptomics Validation. Integration with Xenium 5K spatial transcriptomic data confirmed the model's predictions. Tumor regions showed enrichment of epithelial cancer cells with increased memory T cell infiltration (marked by CD44 and CXCR4). Normal regions were enriched in vascular endothelial cells and alveolar fibroblasts, cell types associated with better overall survival. This spatial validation confirmed that the AI model's differentiation predictions align with the actual cellular composition of different tissue regions.
Enabling Large-Scale Analysis. The core advantage of this approach is scalability. While scRNA-seq is limited to specialized research settings and small sample sizes, H&E slides are produced routinely at every cancer diagnosis center in the world. This AI framework can analyze tumor progression dynamics in thousands of patients using data that already exists in hospital pathology archives, enabling retrospective studies and prospective monitoring at a population scale previously impossible with molecular sequencing methods.
Complementing Pathologist Assessment. The model was validated against expert pathologist annotations and consistently identified tumor regions of interest correctly. Rather than replacing pathologists, the system assists in quantifying subtle spatial and temporal aspects of tumor behavior that are difficult to assess visually. The model can also differentiate necrotic regions from normal tissue and tumor -- a distinction with clinical significance for treatment response assessment.
Known Limitations. The training dataset of 440 annotated ROIs is relatively small, potentially limiting prediction robustness for rare grade categories. The G2/G3 boundary remains the most challenging classification problem due to subjective variation even among pathologists. The fitness score incorporates features from both tumor and non-tumor regions because cell differentiation status reflects similarity to normal tissue -- future work could improve specificity by incorporating cell type segmentation to isolate tumor-specific features.
Future Directions. Expanding the annotated training dataset with a larger variety of tumor grades and pathological conditions would enhance model performance. Incorporating cell type segmentation alongside differentiation status would enable more refined analysis of interactions between tumor, stromal, and immune cells. Application to other cancer types beyond lung adenocarcinoma could validate the generalizability of the pseudotime and fitness score framework.
From Images to Trajectories. This study demonstrates that standard pathology images contain far more information than traditional visual assessment extracts. By applying AI to quantify cell morphological features, spatial organization, and pseudotime, the framework recovers information about tumor evolutionary dynamics that previously required expensive single-cell molecular profiling. This shifts the paradigm from static snapshot analysis of tumors to dynamic trajectory-based understanding.
Prognostic Value Beyond Current Standards. The fitness score, derived purely from H&E images, independently predicts patient survival in three cohorts and improves the concordance index when added to standard clinical staging. This means the method adds genuine prognostic information beyond TNM stage, age, and gender -- the factors currently available in routine clinical practice. Identifying fast-progression patients could inform decisions about adjuvant therapy intensity and follow-up frequency.
Accessibility and Cost-Effectiveness. Unlike molecular biomarkers that require fresh tissue, specialized laboratory equipment, and significant cost, this approach requires only standard formalin-fixed paraffin-embedded tissue slides that are already collected as part of routine care. This makes it applicable in low-resource settings and for retrospective studies of historical patient cohorts, democratizing access to tumor heterogeneity analysis globally.