Current staging misses critical survival-related information. Non-small cell lung cancer (NSCLC) is the leading cause of cancer-related death worldwide. Current patient stratification relies on the UICC TNM staging system, which evaluates tumor size, lymph node involvement, and distant metastasis. While widely used and reproducible, TNM staging frequently fails to predict relapse in early-stage patients, with 21 to 71% of stage I-III patients dying within 5 years of diagnosis.
About half of early-stage patients receiving surgery with curative intent eventually relapse and have poor prognoses. Adjuvant chemotherapy is available but its benefit is uncertain in individual patients. The search for reliable biomarkers to stratify early-stage patients for adjuvant therapy has remained inconclusive even with next-generation sequencing approaches.
The tumor microenvironment (TME) -- the complex ecosystem of immune cells, stromal cells, and extracellular matrix surrounding cancer cells -- has emerged as a key factor in cancer progression and therapeutic response. However, most TME studies have analyzed limited cell types or compared tumor and stroma at an aggregated level, failing to capture the complex spatial relationships between different cell populations that may drive clinical outcomes.
A large-scale multimodal analysis of 1,168 patients. The study enrolled 1,168 NSCLC patients from two major German cancer centers -- Charite Berlin (n=786) and University Hospital Cologne (n=382) -- who underwent surgical tumor resection between 2006 and 2019. The cohort was divided into 673 lung adenocarcinoma (LUAD) and 473 squamous cell carcinoma (LUSC) cases, with tissue microarrays (TMAs) constructed from four representative 1.5 mm tumor cores per patient to enable high-throughput imaging.
A 12-plex multiplex immunofluorescence (mIF) panel characterized immune and tumor cells through expression of CD3, CD4, CD8, CD20, CD56, CD68, CD163, FOXP3, Granzyme B, PD-1, PD-L1, and cytokeratin. This marker combination enabled classification of 43 distinct cell phenotypes. The same sections were also stained with H&E and re-scanned to integrate classical histomorphology with the multiplex data, yielding a database of 57,000 tissue images.
The AI pipeline incorporated 14 distinct deep learning models arranged sequentially. A UNet model with active learning in the loop segmented tissue regions (carcinoma, stroma, necrosis, healthy tissue) achieving macro-averaged F1 of 0.92. An optimized StarDist model detected individual cell nuclei at F1 of 0.91 based on 100,000+ pathologist-annotated cells. Twelve independent ConvNext models -- one per mIF channel -- classified each cell's marker expression status, achieving average F1 of 0.91 across all markers.
Independent classification models were used per mIF channel rather than a joint model to prevent exploiting correlations between channels (Clever Hans effects), which would reduce the biological interpretability of the classifications. In total, 53 million cells were classified and their spatial locations recorded, forming the foundation for downstream niche analysis.
Profound variability in immune cell composition between and within subtypes. Hierarchical clustering of cell type compositions revealed six distinct immune states in LUAD and seven in LUSC, reflecting substantial biological heterogeneity. The full spectrum from immune cell-rich 'hot' to immune cell-poor 'cold' tumors was observed in both subtypes, consistent with patterns described in other cancers.
Cold tumors, characterized by high cancer cell density and very low lymphocyte infiltration (TME type 2; PD-L1 negative, TIL negative), were substantially more common in LUSC than LUAD (14.8% versus 5.6%). These tumors are expected to have poor responses to immune checkpoint therapy and may require alternative approaches such as radiotherapy to recruit lymphocytes.
Tumors expressing PD-L1 alongside high tumor-infiltrating lymphocyte density (TME type 1; PD-L1 positive, TIL positive) were identified in both subtypes and represent the group most likely to respond to checkpoint immunotherapy. Tumors expressing PD-L1 without TIL infiltration (TME type 3) cannot use PD-L1 as a predictive biomarker for immunotherapy response and may benefit from combination regimens.
B cell-rich subtypes were identified in both LUAD and LUSC but were more common in LUAD. Growing evidence supports a role for tumor-infiltrating B cells and plasma cells in cancer biology, and their enrichment in distinct immune state clusters suggests a biologically meaningful subset with potentially distinct prognosis and treatment implications.
From cell counts to spatial neighborhoods. Beyond characterizing overall immune cell composition, the study introduced the concept of 'cell niches' -- defined as the local cellular neighborhood of each cell, computed by counting all cell types within a 34 micrometer radius around each cell. This radius was chosen to capture both directly adjacent cells and second-order neighbors, consistent with previous spatial biology studies.
All 53 million cell neighborhoods were clustered separately for LUAD and LUSC into 10 distinct niches each, using the neighborhood vector (counts of each cell type within the 34 micrometer radius) as the feature. This yielded niches with biologically coherent compositions including carcinoma cell-dominant niches, carcinoma cell-enriched niches with strong immune reaction, lymphocyte-dominant niches, and macrophage-dominant niches.
In LUAD, the 10 niches grouped into five categories: carcinoma cell-dominant (niche 3), carcinoma cell-enriched with low to moderate immune reaction (niches 1, 7, 9), carcinoma niches with strong immune reaction (niches 2, 4, 8), lymphocyte-rich niches (niches 5, 6), and a macrophage-rich niche (niche 10). Similar groupings emerged in LUSC, though with distinct quantitative compositions reflecting the different biology of squamous versus adenocarcinoma histology.
Each niche was further characterized in terms of its biological identity using the underlying cell phenotype distributions. Hot niches included inflamed niches (Inflhigh), strong immune infiltration niches (Leucohigh), lymphocyte-rich niches (Lymphigh I and II), T cell-dominant niches (T cellhigh), and tertiary lymphoid structure-like niches (TLShigh). Cold niches included immune-deserted niches (ImmuNULL), immunosuppressive niches (Immunosuphigh), weak immune infiltration niches (Leucolow), and macrophage-dominated niches (Mhigh).
14% and 47% improvement over standard staging. The UICC8 staging system achieved concordance indices of 0.633 for LUAD and 0.630 for LUSC on the independent Cologne validation cohort (models were trained on the Berlin cohort). These values represent the starting point for comparison.
Adding cell density features to UICC8 staging improved concordance indices to 0.644 for LUAD (8% relative improvement over the chance-level baseline of 0.5) and 0.674 for LUSC (34% relative improvement). However, replacing cell densities with cell niche features combined with UICC8 further improved concordance to 0.665 for LUAD (14% relative improvement) and 0.692 for LUSC (47% relative improvement).
The niche-based models significantly outperformed models using only cell densities (P less than 0.01, Wilcoxon signed-rank test), confirming that the spatial organization of cells within the TME carries prognostic information beyond what is captured by cell counts alone. Kaplan-Meier analyses demonstrated clear survival separation between risk groups assigned by the niche-based models, with patients assigned the lowest niche-based risk scores showing longer survival than UICC8 stage 1 patients.
The risk scores derived from TMA cores showed high correlation with risk scores derived from whole tumor sections in 20 NSCLC cases (Pearson r = 0.979 for LUAD, r = 0.997 for LUSC), with 100% concordance at the level of risk group assignments. This confirms that the TMA-based approach adequately captures tumor-level spatial properties without requiring full-section analysis.
Reclassifying patients missed by current staging. One of the most clinically significant findings was the reclassification of UICC8 stage 1 patients. Under current guidelines, stage 1 patients receive surgery alone without adjuvant chemotherapy. The niche-based model reassigned approximately half of stage 1 patients to a higher risk category (risk score 2 or 3).
These reclassified stage 1 patients showed survival outcomes similar to UICC8 stage 2 patients, who are typically recommended adjuvant chemotherapy. This suggests that a substantial subset of patients currently not receiving adjuvant therapy may in fact benefit from it, representing a potentially large population of undertreated high-risk individuals.
The niche-based risk groups showed no significant association with common clinical confounders including sex, age, ECOG performance status, smoking status, tumor grade, TTF-1 expression, growth pattern, or mutation status of frequently mutated lung cancer genes. This independence from standard clinical variables confirms that niche patterns capture distinct biological information not already encoded in existing clinical data.
The practical path to clinical implementation would involve obtaining biopsy or surgical resection tissue, performing 12-plex mIF imaging, running the AI analysis pipeline to assign cell niches and compute a patient-level risk score (using the maximum risk across TMA cores as the patient-level estimate), and using this risk score alongside UICC8 staging to guide adjuvant therapy decisions.
Hot niches predict better outcomes; cold niches predict worse. Analysis of niche patterns within risk groups confirmed that patients with the best prognosis (risk score 1) had tumors dominated by hot niches including inflamed, strong immune infiltration, and T cell-dominant niches. High densities of tumor-infiltrating lymphocytes including CD3+, CD4+, CD8+, and CD20+ T and B cells are well-established favorable prognostic biomarkers in NSCLC.
Patients in higher risk groups (risk scores 2 and 3) showed progressive accumulation of cold niches including immune-deserted niches, immunosuppressive niches, weak immune infiltration niches, and macrophage-dominated niches. Macrophage-rich niches are consistently associated with reduced survival across cancer types, and the identification of specific cold niche patterns linked to high risk adds mechanistic specificity beyond generic 'cold tumor' classification.
Subtype-specific differences in niche prognostic associations were observed. Tertiary lymphoid structure-like niches were associated with favorable prognosis in LUSC (risk score 1) but with higher risk groups in LUAD, highlighting that niche patterns must be interpreted in the context of tumor histology. This underscores the importance of separate LUAD and LUSC analyses rather than treating NSCLC as a single entity.
A paradigm shift in NSCLC risk stratification. This study demonstrates that AI-powered multiplex imaging analysis of spatial cell niches in the tumor microenvironment provides prognostic information substantially beyond current TNM staging, with the largest improvements seen in LUSC where the spatial immune architecture appears particularly informative relative to simple staging.
The identification of undertreated high-risk stage 1 patients as a major clinical application demonstrates the practical value of the approach. Current guidelines withhold adjuvant therapy from stage 1 patients, but the niche-based model identifies a subset with stage 2-equivalent prognosis who may benefit -- a question that could now be tested in prospective trials stratified by niche-based risk scores.
The 12-plex mIF approach used in this study is technically demanding and currently limited to specialized centers, but the authors note that the robustness of a standardized multiplex panel makes broader diagnostic implementation realistic in the near future as mIF technology becomes more widely available.
Future directions include validation in broader patient populations beyond German academic centers, incorporation of cancer-associated fibroblast subtypes not captured in the current mIF panel, integration with multi-omics data to elucidate the mechanisms of niche formation, and experimental perturbation of identified cell niches to establish causal relationships between niche patterns and therapeutic response.