A spatial immune scoring system based on the tumor microenvironment features improves prognostic stratification in lung adenocarcinoma Authors: Wei Luo, Jia Jia, Qi Wang, Xiaoqing Li, Fang Yang, Yongtian Zhao, Zilu Li, Lin Li, Xiaodan Wang, Mingyao Meng, Haiyan Guo, Wanli Zheng, Liqiong Wang, Dingbiao Li, Zongliu Hou, Yangfan Guo Published: 2026 ABSTRACT Highlights•Lung Cancer Immune Score quantifies the balance between host immunity and tumor aggressiveness.•High Lung Cancer Immune Score independently predicts better overall survival.•Lung Cancer Immune Score refines Stage II and III LUAD prognosis beyond TNM staging. Introduction Lung adenocarcinoma (LUAD) is the most common subtype of non-small cell lung cancer (NSCLC) and a leading cause of cancer-related mortality worldwide [1]. Despite recent progress in multimodal therapies, patient prognosis remains highly variable. The five-year survival rates decrease significantly from approximately 70% in localized disease to <10% in metastatic cases [2,3]. This variability is largely attributed to the profound biological heterogeneity of the tumor and the limitations of conventional staging systems. In current clinical practice, risk stratification depends mainly on the TNM classification and molecular profiling (e.g., EGFR, ALK, or KRAS mutations) [4,5]. However, TNM staging primarily evaluates the anatomical tumor burden and cannot fully reflect biological aggressiveness [6,7]. Furthermore, although molecular subtyping guides targeted therapies, genomic features are insufficient to determine clinical outcomes, mainly because they do not fully explain the complexity of host antitumor immunity [8]. Consequently, patients presenting with the same TNM stages and mutation profiles often experience divergent survival outcomes. This clinical variability highlights the critical role of the tumor immune microenvironment (TME), a complex spatial ecosystem that significantly influences tumor progression, immune surveillance, and therapeutic resistance [9]. Accumulating evidence indicates that patient prognosis depends not only on the overall abundance of tumor-infiltrating immune cells, but also on their functional states and spatial distribution within the tumor architecture [[10], [11], [12]]. In LUAD, the TME exhibits a wide spectrum of immunological profiles, ranging from immune-inflamed (“hot”) phenotypes, characterized by effective antigen presentation, tertiary lymphoid structure (TLS) formation and robust cytotoxic T cell infiltration, to immune-excluded or immune-desert (“cold”) phenotypes that reflect tumor immune evasion [13]. Within this spatial context, different immune cell populations exert non-redundant functions. For instance, CD8⁺ cytotoxic T cells mediate tumor cell killing, M1-like tumor-associated macrophages promote pro-inflammatory and anti-tumor responses, B cells contribute to antigen presentation and TLS formation, and professional antigen-presenting cells orchestrate T cell priming and activation, among others. Consequently, shifting from conventional bulk tissue analysis to spatially resolved immune profiling is essential to supplement established clinical staging and elucidate the biological determinants of patient prognosis. Immune scoring systems have demonstrated significant clinical utility in solid tumors, particularly in colorectal carcinoma, where spatial immune metrics often outperform traditional anatomical staging in predicting disease recurrence [14,15]. However, LUAD features a distinct immune microenvironment and complex patterns of immune evasion, highlighting the need for a lung-specific prognostic model based on relevant immunological characteristics [[16], [17], [18]]. To address this, we developed the Lung Cancer Immune Score (LCIS), a spatial prognostic model designed to integrate complex TME architecture into a clinically applicable tool. Using multiplex immunofluorescence and machine learning-assisted spatial profiling, the LCIS integrates four biologically relevant markers: CD68⁺HLA-DR⁺ antigen-presenting macrophages, CD20⁺ B cells, PD-1⁺ immune cells, and Ki67⁺ proliferating tumor cells. The model quantifies their respective densities across the tumor core and the invasive margin. Conceptually, the LCIS evaluates the balance between tumor proliferation (Ki67⁺) and host anti-tumor immunity, represented by antigen presentation (CD68⁺HLA-DR⁺), B cell infiltration (CD20⁺), and immune activation/exhaustion status (PD-1⁺). Here, we demonstrate that this spatial scoring system significantly improves prognostic stratification compared to conventional parameters, providing a valuable biological rationale to guide personalized therapeutic strategies in LUAD. Result Cohort baseline characteristics The retrospective cohort included 63 patients with LUAD who underwent thoracic surgical procedures between 2015 and 2021 and were ultimately staged as IB-IV according to the American Joint Committee on Cancer (AJCC) 8th edition TNM staging system. Demographic parameters showed a median age of 59.87 years (±7.8 years) with a balanced gender distribution (male: 46% [29/63], female: 54% [34/63]). A smoking history was present in 36.5% (23/63) of patients. Patients were distributed across stages IB to IV, with 41.3% (26/63) classified as stage IB. (Table S1). Postoperative recurrence or metastasis occurred in 25 patients, including skeletal (9 cases), contralateral pulmonary (8 cases), cerebral (6 cases), and mediastinal (5 cases). 22 patients died of disease-related complications during the study period. Adjuvant chemotherapy (pemetrexed with platinum regimens) was applied to 87.3% patients (55/63), with 7 patients (11/63) receiving metastasis-directed radiotherapy. Targeted therapy and checkpoint immunotherapy were used in 19 patients (30.2%) and 9 patients (14.3%) respectively. Based on baseline demographic and clinical characteristics, we performed univariate Cox regression analysis to evaluate the prognostic value of clinical parameters. As summarized in Fig. 1d, the TNM stage emerged as the only significant predictor of survival (HR = 1.90, p = 0.002), underscoring its established role in risk stratification. In contrast, conventional hematologic and demographic markers—such as age(HR = 1.00, p = 0.48), smoking history(HR = 0.86, p = 0.75), NLR(HR = 1.10, p = 0.07), WBC(HR = 1.30, p = 0.1)—showed no statistically significant associations with survival outcomes. Similarly, therapeutic interventions like targeted therapy (HR = 1.70, p = 0.270) and immunotherapy (HR = 0.94, p = 0.916) failed to independently stratify prognosis, likely due to the cohort’s limited sample size and heterogeneous molecular profiles. Consistently, conventional molecular subtyping markers (EGFR, KRAS, and TP53) also lacked prognostic significance in both our discovery and the TCGA cohorts (Figure S1). These findings highlight the inadequacy of routine clinical parameters in capturing the biological complexity of LUAD prognosis. While TNM staging provides anatomic stratification, it neglects critical immunological and molecular determinants of tumor behavior. This gap underscores the imperative for integrating TME derived biomarkers, such as spatially resolved immune cell infiltration patterns, to refine prognostic precision and guide personalized therapeutic strategies.Fig. 1Study design and spatial tumor microenvironment characteristics associated with clinical prognosis.a Cohort selection flowchart and overall workflow of spatial tumor microenvironment profiling.b Kaplan-Meier survival curve stratified by TNM stage.c Visualization of tissue classification results using the HALO algorithm. The left panel shows a representative whole-slide image (WSI). The middle and right panels show magnified views of the carcinoma region (CR) and the carcinoma-para-carcinoma junction, respectively. Color labels indicate automated region annotation: carcinoma region (CR), para-carcinoma region (PCR), carcinoma-para-carcinoma junction, stromal region (SR), and tumor region (TR).d Univariate Cox regression analysis showing HR and 95% confidence intervals for clinical characteristics in relation to overall survival.e Heatmap showing median-normalized cell densities (Z-score across all samples) for immune and tumor cell types in different tumor regions, Immune phenotypes were identified using mIHC with four functional panels: Immune Checkpoint & Cytotoxicity (ICP & Cyto), T cell Differentiation & Function (TDiff & Func), Macrophage Polarization & Regulation (Mφ Polar & Reg), and Lymphocyte Exhaustion & Memory (LExh & Mem).f Cell density distribution (cells/mm²) of antigen-presenting cells, PD-L1⁺ cells, proliferating tumor cells, and CD20⁺ cells across different tumor regions.* mark significant inter-regional differences (Wilcoxon rank-sum test, * p < 0.05, ** p < 0.01, *** p < 0.001).Fig 1 dummy alt text Study design and spatial tumor microenvironment characteristics associated with clinical prognosis. a Cohort selection flowchart and overall workflow of spatial tumor microenvironment profiling. b Kaplan-Meier survival curve stratified by TNM stage. c Visualization of tissue classification results using the HALO algorithm. The left panel shows a representative whole-slide image (WSI). The middle and right panels show magnified views of the carcinoma region (CR) and the carcinoma-para-carcinoma junction, respectively. Color labels indicate automated region annotation: carcinoma region (CR), para-carcinoma region (PCR), carcinoma-para-carcinoma junction, stromal region (SR), and tumor region (TR). d Univariate Cox regression analysis showing HR and 95% confidence intervals for clinical characteristics in relation to overall survival. e Heatmap showing median-normalized cell densities (Z-score across all samples) for immune and tumor cell types in different tumor regions, Immune phenotypes were identified using mIHC with four functional panels: Immune Checkpoint & Cytotoxicity (ICP & Cyto), T cell Differentiation & Function (TDiff & Func), Macrophage Polarization & Regulation (Mφ Polar & Reg), and Lymphocyte Exhaustion & Memory (LExh & Mem). f Cell density distribution (cells/mm²) of antigen-presenting cells, PD-L1⁺ cells, proliferating tumor cells, and CD20⁺ cells across different tumor regions. * mark significant inter-regional differences (Wilcoxon rank-sum test, * p < 0.05, ** p < 0.01, *** p < 0.001). Spatial compartmentalization of the immune microenvironment Histopathological segmentation of the tumor architecture was performed through initial H&E staining on all tissue specimens, enabling morphologically-guided annotation of five distinct analytical compartments: whole-slide imaging, carcinoma, para-carcinoma, tumor and stromal regions. The HALO Tissue Classifier algorithm was trained on DAPI and CK/SOX10 channels to automatically segment tumor regions (marked in red) and stromal regions (marked in green). The algorithm further quantified the expression levels of distinct immune cell subsets within the tumor and stromal compartments (Fig. 1c). To characterize immune microenvironment heterogeneity, we implemented a comprehensive multiplex immunohistochemistry (mIHC) using four 7-marker panels (including DAPI nuclear counterstain) in 63 treatment-naïve lung adenocarcinoma specimens. Tumor-stromal differentiation was systematically performed using Pan-CK-based deep learning segmentation algorithms, enabling compartment-specific analysis across all anatomically defined regions. Spatial distribution analysis revealed compartmentalized immune cell enrichment patterns in LUAD tissues (Fig. 1d-e). Proliferating tumor cells (Pan-CK⁺) dominated carcinoma regions (median density: 315.83 vs. para-carcinoma 8.69 cells/mm², p < 0.001), where exhausted T cells (TIM3⁺, LAG3⁺) were also predominantly localized. In contrast, stromal regions exhibited significant accumulation of PD-1⁺ immune cells (262.61 vs. tumor 85.11 cells/mm², p = 0.004), immunosuppressive myeloid subsets (e.g., CD163⁺ macrophages), and cytotoxic T cells (Table S3). This spatial dichotomy suggests functional specialization: carcinoma nests may foster T cell exhaustion and tumor proliferation, while stromal zones harbor both immunosuppressive interactions (via PD-1⁺ cells and macrophages) and cytotoxic T cell reservoirs. CD20⁺ B cells were predominantly localized in the tumor (4812.34 vs. para-carcinoma 2998.46 cells/mm², p < 0.001), suggesting compartmentalized humoral immune activity. Technical limitations in the LExh & Mem Panel (absence of Pan-CK detection) precluded precise tumor-stroma discrimination, restricting comparative analyses to broader tumor versus stromal regions. Collectively, these findings highlight dynamic crosstalk between spatially segregated immune niches, potentially driving therapeutic resistance through microenvironmental partitioning. Prognostic survival analysis based on immune characteristics To evaluate the prognostic relevance of tumor-infiltrating immune cells (TIICs), we stratified patients into high/low infiltration groups based on median TIIC densities within distinct tissue compartments (carcinoma, Para-carcinoma, tumor, stromal region, and WSI) and performed Kaplan-Meier survival analysis. Using a more lenient statistical threshold (p ≤ 0.1) to account for the exploratory nature of this cohort, we identified spatially resolved immune subsets with suggestive associations to overall survival (Fig. 2). In the stromal region, elevated densities of antigen-presenting cells (p = 0.041) and CD68+&HLA-DR+ myeloid cells (p = 0.041) correlated with improved survival. In the carcinoma region, increased infiltration of antigen-presenting cells (APCs) (p = 0.031), CD68+&HLA-DR+ myeloid cells (p = 0.038) predicted favorable outcomes. Whole-slide analyses further revealed a trend between CD20+ B cell infiltration and prolonged OS (p = 0.072), whereas higher PD-1⁺ cell density in the WSI tumor region showed a trend toward poorer OS.Fig. 2Kaplan-Meier survival curves showing the prognostic significance of nine selected biomarkers when analyzed within specific TME compartments. Each panel (a-i) shows a KM curve for a different biomarker-compartment combination. Abbreviations indicate analytical compartments: C, carcinoma region; CT, tumor region within the carcinoma region; CS, stromal region within the carcinoma region; WSIT, tumor region within the whole-slide image.Fig 2 dummy alt text Kaplan-Meier survival curves showing the prognostic significance of nine selected biomarkers when analyzed within specific TME compartments. Each panel (a-i) shows a KM curve for a different biomarker-compartment combination. Abbreviations indicate analytical compartments: C, carcinoma region; CT, tumor region within the carcinoma region; CS, stromal region within the carcinoma region; WSIT, tumor region within the whole-slide image. From these findings, nine biomarker-compartment features demonstrating consistent survival trends and biological relevance were prioritized as candidate prognostic markers:1)Proliferating cells (Ki67+) in the carcinoma (Tumor Region).2)Proliferating tumor cells (Ki67+&Pan-CK+) across whole-slide images.3)CD68+&HLA-DR+ myeloid cells in the carcinoma.4)CD68+&HLA-DR+ myeloid cells in the carcinoma (Tumor Region).5)CD68+&HLA-DR+ myeloid cells in the carcinoma (Stromal Region).6)Antigen-presenting cells (HLA-DR+) in the carcinoma (Stromal Region).7)Antigen-presenting cells in the carcinoma (Tumor Region).8)CD20+ B cells across whole-slide images.9)PD-1+ cells in the whole-slide images (Tumor Region). 1)Proliferating cells (Ki67+) in the carcinoma (Tumor Region). Proliferating cells (Ki67+) in the carcinoma (Tumor Region). 2)Proliferating tumor cells (Ki67+&Pan-CK+) across whole-slide images. Proliferating tumor cells (Ki67+&Pan-CK+) across whole-slide images. 3)CD68+&HLA-DR+ myeloid cells in the carcinoma. CD68+&HLA-DR+ myeloid cells in the carcinoma. 4)CD68+&HLA-DR+ myeloid cells in the carcinoma (Tumor Region). CD68+&HLA-DR+ myeloid cells in the carcinoma (Tumor Region). 5)CD68+&HLA-DR+ myeloid cells in the carcinoma (Stromal Region). CD68+&HLA-DR+ myeloid cells in the carcinoma (Stromal Region). 6)Antigen-presenting cells (HLA-DR+) in the carcinoma (Stromal Region). Antigen-presenting cells (HLA-DR+) in the carcinoma (Stromal Region). 7)Antigen-presenting cells in the carcinoma (Tumor Region). Antigen-presenting cells in the carcinoma (Tumor Region). 8)CD20+ B cells across whole-slide images. CD20+ B cells across whole-slide images. 9)PD-1+ cells in the whole-slide images (Tumor Region). PD-1+ cells in the whole-slide images (Tumor Region). This selective approach balances statistical significance with biological plausibility, avoiding premature exclusion of candidates that may synergize in combinatorial models. Development and evaluation of the LCIS To develop a robust prognostic framework tailored for LUAD, we constructed the LCIS based on spatially resolved immune metrics. We first evaluated a T cell-centric scoring system adapted from colorectal cancer[15,19], which was based on the densities of CD3⁺ and CD8⁺ T cells. Densities were quantified across the tumor samples and stratified into five levels (I0 to I4). This system failed to stratify patient outcomes in our cohort (OS p = 0.91, PFS p = 0.92), suggesting that LUAD has a distinct immunogenic landscape (Fig. 3a). We then developed and compared two alternative scoring frameworks based on the same candidate marker set: LCIS1, based on percentile thresholds (HR = 0.71, p = 0.002, Table 1), and LCIS2, utilizing a spatial fold-change ratio approach (HR = 0.91, p = 0.003, Table 1). Although both models demonstrated improved prognostic accuracy, we further optimized the system into LCIS3 to improve its clinical utility and generalizability (see Supplementary Appendix 1 and Figure S2 for detailed model development workflow).Fig. 3LCIS3 identifies prognostic subgroups of LUAD associated with distinct spatial immune microenvironment features.a Survival probability stratified by literature-based immune scores.b Survival analysis of High vs. Low LCIS3 groups.c Spatial Immunofluorescence in LCIS subtypes.d Hazard ratio associations between LCIS3 and clinical characteristics.e Original cell abundance differences between High vs. Low LCIS groups.f Carcinoma / Para-carcinoma cell abundance differences with log2 normalization of between High vs. Low LCIS groups; analyzed by Wilcoxon rank-sum test; * p < 0.05, ** p < 0.01, *** p < 0.001, **** p < 0.0001.Fig 3 dummy alt textTable 1Predictive performance of the three prognostic models in the discovery cohort.Table 1 dummy alt textModelCox P valueHR (95% CI)C-index3-year AUC5-year AUCLCIS10.0020.71 (0.57–0.88)0.7370.800.77LCIS20.0030.91 (0.85–0.97)0.7040.770.71LCIS30.0100.83 (0.71–0.97)0.7000.750.65 LCIS3 identifies prognostic subgroups of LUAD associated with distinct spatial immune microenvironment features. a Survival probability stratified by literature-based immune scores. b Survival analysis of High vs. Low LCIS3 groups. c Spatial Immunofluorescence in LCIS subtypes. d Hazard ratio associations between LCIS3 and clinical characteristics. e Original cell abundance differences between High vs. Low LCIS groups. f Carcinoma / Para-carcinoma cell abundance differences with log2 normalization of between High vs. Low LCIS groups; analyzed by Wilcoxon rank-sum test; * p < 0.05, ** p < 0.01, *** p < 0.001, **** p < 0.0001. Table 1Predictive performance of the three prognostic models in the discovery cohort.Table 1 dummy alt textModelCox P valueHR (95% CI)C-index3-year AUC5-year AUCLCIS10.0020.71 (0.57–0.88)0.7370.800.77LCIS20.0030.91 (0.85–0.97)0.7040.770.71LCIS30.0100.83 (0.71–0.97)0.7000.750.65 Predictive performance of the three prognostic models in the discovery cohort. LCIS3 was derived from the same candidate marker set as LCIS1 and LCIS2 but was simplified to four biologically interpretable markers: CD68⁺HLA-DR⁺ myeloid cells, CD20⁺ B cells, Ki67⁺ proliferating tumor cells, and PD-1⁺ immune cells. For each retained marker, a spatial enrichment score was calculated as the log2-transformed density ratio between the carcinoma and para-carcinoma regions. CD68⁺HLA-DR⁺ myeloid cells and CD20⁺ B cells were assigned positive weights, whereas Ki67⁺ tumor cells and PD-1⁺ immune cells were assigned negative weights. The final LCIS3 score integrated these weighted spatial enrichment components, with the full formula, cutoff determination, and model development workflow provided in the Supplementary Appendix and Figure S2. Kaplan-Meier analysis revealed significantly improved overall survival in the high-LCIS3 group, and multivariable Cox regression confirmed that LCIS3 was independently associated with overall survival (HR = 0.80, p = 0.02; Fig. 3d). Stratification by the LCIS3 identified different TME profiles for each risk group (Fig. 3c). We observed that the low-score group had a higher density of proliferating tumor cells (p < 0.001, Fig. 3e), while the high-score group was characterized by the accumulation of CD68⁺HLA-DR⁺ myeloid cells (p < 0.01, Fig. 3f). Notably, the LCIS3 model showed better discriminatory power compared to single-marker analyses, and spatial validation confirmed the enrichment of these four cell types in tumor cores compared to invasive margins in individual samples. Additional sensitivity analyses were performed to assess the potential influence of stage heterogeneity and treatment-related confounding. After excluding stage IV patients, the effect direction of LCIS3 remained consistent, although the strength of association was attenuated and no longer reached statistical significance. Similarly, after excluding patients who received targeted therapy or immunotherapy, LCIS3 still showed a significant survival separation in Kaplan-Meier analysis (p = 0.019) and remained significantly associated with overall survival in univariable analysis (p = 0.035). (Table S4 and Figure S3). Validation of LCIS in independent cohort To assess the generalizability of the LCIS3, we validated it in an independent cohort comprising 90 treatment-naïve NSCLC patients. Tissue microarray (TMA) sections were analyzed using mIHC with protocol-matched antibody panels to quantify the tumor core densities of four predefined subtypes: CD68+&HLA-DR+myeloid cells, proliferating tumor cells, PD-1+ cells and CD20+ B cells. Patients were stratified into high- and low-LCIS3 groups using the previously established threshold (−5.0) derived from the discovery cohort using maximally selected log-rank statistics. KM analysis demonstrated a significant prolonged OS in the high-LCIS3 group compared with low-scoring patients (p = 0.017, Fig. 4c). These results validate the clinical utility of spatially resolved immune profiling and suggest LCIS3 may enhance risk stratification beyond conventional clinicopathological parameters.Fig. 4Independent validation of LCIS3 stratification.a Representative spatial immunofluorescence images from an independent validation cohort. Image display staining for DAPI, CD20, HLA-DR & CD68, PAN—CK & Ki67 and PD-1.b Key cell subtype abundance differences between High vs. Low LCIS groups, analyzed by Wilcoxon rank-sum test; * p < 0.05, ** p < 0.01, *** p < 0.001.c Kaplan-Meier survival curves comparing OS between LCIS3-high and LCIS3-low samples.Fig 4 dummy alt text Independent validation of LCIS3 stratification. a Representative spatial immunofluorescence images from an independent validation cohort. Image display staining for DAPI, CD20, HLA-DR & CD68, PAN—CK & Ki67 and PD-1. b Key cell subtype abundance differences between High vs. Low LCIS groups, analyzed by Wilcoxon rank-sum test; * p < 0.05, ** p < 0.01, *** p < 0.001. c Kaplan-Meier survival curves comparing OS between LCIS3-high and LCIS3-low samples. These findings provide preliminary external support for the prognostic stratification ability of LCIS3, suggesting a certain degree of model stability and potential clinical value. LCIS enhances prognostic stratification in intermediate-stage LUAD To further explore LCIS3 as a complementary prognostic tool to TNM staging, particularly for finer risk stratification within established clinical stages, we combined the discovery and validation cohorts to enhance statistical power for subgroup analysis. We specifically focused on patients with intermediate-stage (II and III) disease. In the pooled cohort of stage II and III patients, LCIS3 demonstrated slightly superior prognostic stratification (p = 0.043) compared to TNM staging alone (p = 0.05), which distinguishes between stage II and III survival (Fig. 5a, 5b). In TNM stage II patients, a high LCIS3 score was significantly associated with superior overall survival (OS) (p = 0.05, Fig. 5c). A similar, though non-significant, trend towards better OS with high LCIS3 was observed in TNM stage III patients (p = 0.071, Fig. 5d), with the limited sample size possibly contributing to the lack of statistical significance. Collectively, these findings underscore the potential of LCIS3 to delineate distinct prognostic subgroups within the same TNM stage, serving as a valuable adjunct to conventional staging for more personalized risk assessment and potentially guiding therapeutic decisions.Fig. 5LCIS3 enhances prognostic stratification in intermediate-stage LUAD.a Kaplan-Meier survival curves comparing OS between stage II and stage III samples in the pooled cohort of TNM II and III stage patients.b Kaplan-Meier survival curves comparing OS between LCIS3-high and LCIS3-low samples in the pooled cohort of TNM II and III stage patients.c, d Kaplan-Meier survival curves comparing OS between LCIS3-high and LCIS3-low samples within TNM stage II (c) and stage III (d) patients, respectively.e Schematic model illustrating the proposed anti-tumor immune cascade underlying a high LCIS3 score.Fig 5 dummy alt text LCIS3 enhances prognostic stratification in intermediate-stage LUAD. a Kaplan-Meier survival curves comparing OS between stage II and stage III samples in the pooled cohort of TNM II and III stage patients. b Kaplan-Meier survival curves comparing OS between LCIS3-high and LCIS3-low samples in the pooled cohort of TNM II and III stage patients. c, d Kaplan-Meier survival curves comparing OS between LCIS3-high and LCIS3-low samples within TNM stage II (c) and stage III (d) patients, respectively. e Schematic model illustrating the proposed anti-tumor immune cascade underlying a high LCIS3 score. Discussion In this study, we developed a spatially resolved immune scoring system, LCIS, to quantify tumor core to boundary immune density. By integrating four key biomarkers (CD68+ & HLA-DR+ myeloid cells, PD-1⁺ cells, CD20⁺ B cells, and proliferating tumor cells), LCIS3 emerged as an independent prognostic indicator for LUAD, with patients with high scores showing significantly improved clinical outcomes. This spatial mapping framework not only enhances risk stratification, particularly for patients with indeterminate risk according to conventional TNM stage but also offers valuable insights for elucidating the prognostic significance of spatial immune subsets. TME acts as a crucial ecosystem in oncogenesis, where spatially distributed immune cells dynamically regulate anti-tumor immune responses. Nowadays, PD-1/PD-L1 checkpoint blockade has significantly transformed first-line management of advanced NSCLC. However, significant interpatient heterogeneity in therapeutic responses persists, driven largely by the compositional diversity of the TME [20]. This biological complexity highlights the prognostic value of immune infiltration patterns and underscores the utility of multiplex immunofluorescence-based TME profiling as a approach for identifying novel prognostic biomarkers[21,22]. Our analysis revealed distinct TME landscapes. Specifically, we observed that immune cells predominantly accumulate in tumor core regions rather than at tumor margins. Both lymphoid cells (including helper T cells, cytotoxic T cells, and memory T cells) and myeloid cells (including various macrophage populations and B cells) showed a clear preference for localization within the tumor core. These results align with the refined immune cell atlases of early adenocarcinoma, demonstrating that significant alterations in lymphoid and myeloid lineages emerge early in oncogenesis and continue across clinical stages [[23], [24], [25]]. While bulk studies have largely associated diverse immune infiltrates, including M2 macrophages, helper T cells, CD8+ T cells, NK cells, M1 macrophages, B cells and dendritic cells, with favorable tumor prognosis[[26], [27], [28]], our multivariate Cox regression analyses of individual immune subsets failed to identify independent prognostic factors. However, the spatial quantification revealed a clinically significant pattern. Specifically, patients exhibiting elevated APC densities and M1 macrophage enrichment within the TME demonstrated superior OS and PFS. Consistent with our findings, recent studies reported high M1 macrophages infiltration was associated with improved prognosis, while elevated CD68+&CD163+ cell density correlated with poorer OS in lung cancer[[27], [28], [29]]. This prognostic advantage likely related to M1 macrophages orchestrating cytotoxic T lymphocytes activation via antigen presentation or pro-inflammatory cytokines secretion, resulting in T cell proliferation and IFN-γ-mediated tumor cell cytotoxicity[30]. The four immunologically relevant markers integrated into LCIS3 may collectively reflect the dynamic balance between tumor proliferative pressure and host immune response within the TME (Fig. 6).Fig. 6Proposed immune-proliferative balance captured by LCIS3 in the lung adenocarcinoma tumor microenvironment.Fig 6 dummy alt text Proposed immune-proliferative balance captured by LCIS3 in the lung adenocarcinoma tumor microenvironment. In this framework, Ki67+ tumor cells represent proliferative burden and biological aggressiveness[31,32]. Increased proliferative activity is often accompanied by accelerated cell turnover, genomic instability, and antigen release, which may influence tumor immune visibility and subsequent immune recognition[22]. In parallel with tumor-side proliferative pressure, CD68+HLA-DR+ myeloid cells and CD20+ B cells may reflect the immune-response side of the TME, including local antigen presentation, immune-cell coordination, and the organization of anti-tumor immune niches[33,34]. Within this immune context, PD-1+ cells may indicate immune cells that have encountered tumor antigens and are involved in ongoing local immune responses; however, sustained PD-1 expression may also reflect checkpoint-mediated functional restraint under chronic antigen stimulation[[35], [36], [37]]. Taken together, these four markers may capture complementary biological dimensions of the LUAD microenvironment: tumor proliferation, antigen-presenting myeloid activity, B-cell involvement, and PD-1-related immune engagement. A high LCIS3 score may therefore reflect a spatial context in which host immune response is relatively preserved in relation to tumor proliferative pressure. Conversely, a low LCIS3 score may indicate a microenvironment in which proliferative tumor pressure predominates over organized anti-tumor immune activity. The LCIS establishes a spatial quantification assessment framework capturing critical functional dimensions of the TME in LUAD. By quantifying spatial immune patterns, including M1 macrophage polarization gradients and PD-1⁺ cell redistribution dynamics, LCIS identifies prognostically decisive TME states distinct from anatomic tumor spread metrics in TNM staging. For instance, high LCIS scores correlate with inflamed (hot) tumor phenotype characterized by enhanced antigen presentation and macrophage activation, which is associated with improved immunotherapy responses [38,39]. Spatial resolution of immune evasion patterns (e.g., immune-evasive pattern at tumor-stromal interfaces) may guide rational combination therapies targeting specific TME compartments. Integration of LCIS with TNM staging could advance precision stratification and personalized treatment paradigms in LUAD. Despite the promising prognostic stratification potential of LCIS, several limitations should be acknowledged. First, this was a single-center retrospective study with a relatively limited discovery cohort size and a small number of events, which may affect the stability of model estimation. Although we included an independent validation cohort and externally assessed the prognostic stratification ability of LCIS3, this validation should still be considered preliminary, and the clinical applicability and generalizability of LCIS require further confirmation in larger, multicenter, prospective cohorts. Stage heterogeneity and treatment-related confounding may still have influenced survival outcomes. Although sensitivity analyses supported the overall direction of LCIS3, these restricted analyses included fewer patients and events; therefore, the subgroup findings should be interpreted cautiously and require confirmation in larger clinically homogeneous cohorts. In addition, the technical complexity in mIHC workflow and AI-driven digital pathology analysis require automated streamlining protocols for wide clinical application. Furthermore, the functional causation between LCIS-derived immune indicators and therapeutic outcomes requires deeper exploration via high-throughput spatial omics and 3D tumor profiling. Addressing these gaps will determine clinical translatability under real-world heterogeneity. In conclusion, this study emphasizes the important role of spatial TME profiling in refining prognostic stratification for LUAD. By integrating immune cell distribution and tumor characteristics, LCIS enhances prognosis prediction and supports more personalized treatment strategies. Methods Patients Between 2015–2021, 349 patients with pathologically confirmed lung adenocarcinoma who underwent thoracic surgical procedures at Yan'an Hospital Affiliated to Kunming Medical University were screened. The inclusion criteria comprised: (1) patients with pathological confirmation of lung adenocarcinoma who had undergone surgical treatment; (2) tumor tissue samples encompassing both the carcinoma and para-carcinoma regions, with a minimum tumor area of 3 mm²; (3) complete clinical follow-up data. Exclusion criteria included: (1) patients presenting with pulmonary metastatic tumors; (2) patients pathological diagnosis without adenocarcinomas; (3) patients with insufficient tumor tissue samples or tumor masses; (4) patients who were lost to follow-up; and (5) patients lacking complete clinical information. Following these criteria, 63 patients were eligible and enrolled for further analysis. (Fig. 1a) Eight patients who were ultimately classified as stage IV underwent exploratory thoracic surgery because they initially presented with pulmonary occupying lesions and were not known to have unequivocally unresectable metastatic disease before surgery. Data collection Patient baseline data were collected including gender, age, smoking history, treatment modality, and TNM stage. Laboratory markers included hemoglobin (HB), white blood cell count, neutrophil count, lymphocyte count, neutrophil-to-lymphocyte ratio (NLR), platelet count, platelet-to-lymphocyte ratio (PLR), serum sodium (Na), albumin, and lactate dehydrogenase (LDH). NLR was calculated by dividing the absolute neutrophil count by the absolute lymphocyte count, while PLR was determined by dividing the platelet count by the lymphocyte count. Information on targeted therapies, immunotherapy, and disease progression or mortality was obtained through telephone follow-up, concluding in July 2023. The study endpoints were disease progression or death. PFS was defined as the time from surgery to the first documented disease progression (tumor recurrence, metastasis, or death). Overall survival (OS) was defined as the time from surgery to death. All patients completed follow-up, with a minimum follow-up was 21 months. Immune cell marker selection and multiplex immunofluorescence panel design To systematically characterize tumor immune infiltration, we selected canonical immune markers through a comprehensive literature review, including CD68 for macrophages, CD3 for total T cells, CD8 for cytotoxic T lymphocytes, CD20 for B lymphocytes, CD56 for natural killer cells, and CD45RO for memory T cells. These markers were further complemented by functional markers capturing dynamic immune states: immune evasion pathways (PD-1 and PD-L1), cellular activation status (HLA-DR), proliferative activity (Ki67), and exhaustion/dysfunction signatures (TIM3, LAG3). The 19 selected markers were strategically grouped into four multiplex immunofluorescence panels based on three interconnected principles: functional synergy (e.g., pairing lineage markers with their associated functional states), spatial co-localization hypotheses (e.g., tumor-immune interface interactions), and spectral signal compatibility. This multi-step sequential staining protocol—combining spectral imaging with iterative antibody stripping—enabled precise quantification of immune cell densities, spatial distributions, and functional states within the tumor microenvironment. Critical co-expression patterns were analyzed to delineate functional immune subtypes, including macrophage polarization states such as CD68+&PD-L1+ immune-evasive niches and CD68+&CD163+/CD206+ M2-like TAMs. T cell functional heterogeneity was resolved through phenotypes like CD8+&PD-1+ exhausted cytotoxic lymphocytes, CD8+&GZMB+ activated effectors, and CD3+&CD45RO+ memory T cell reservoirs. Tumor-immune interplay was decoded through spatial relationships, such as Ki67+PanCK+ proliferating tumor cell clusters juxtaposed with TIM3+&CD20+ exhausted B cell infiltrates. By integrating single-marker quantification with multi-marker co-expression logic, this approach bridges cellular identity, functional states, and spatial context to define immune-driven prognostic subgroups. Quantitative spatial profiling of multiplex immunofluorescence data Quantitative image analysis was performed on the HALO® Digital Pathology Platform (V3.5.3577; Indica Labs, Albuquerque, NM, USA). Initial histopathological annotation defined regions of interest (ROIs), including carcinoma, para-carcinoma, tumor, and stromal regions through co-registration with H&E reference sections, excluding artifacts and necrotic areas. Cell segmentation and phenotyping were executed using the HALO Highplex FL algorithm (v4.2.14), with nuclear detection sensitivity and subcellular localization-based intensity thresholds (nuclear, cytoplasmic, or membrane) for positive cell identification. Nuclear segmentation was trained via the HALO AI Nuclei Seg (FL) deep learning algorithm on 20 manually annotated 1 × 1 mm² regions from five mIHC samples (DAPI channel), with nuclear size constrained to 11.3–571.7 μm² and minimum nuclear roundness set to 0.3 to exclude irregular nuclei. Cytoplasmic boundaries were defined as a 3 μm extension from nuclear edges, and marker-specific intensity thresholds were applied based on subcellular staining localization. This method provided subcellular spatial resolution for immune infiltrate quantification across all 19 markers. Statistical analysis Statistical analysis was performed using SPSS 17.0 and R 4.4.2. Continuous variables with a normally distribution are expressed as mean ± SD, while non-normally distributed data were presented as median (IQR). Categorical associations were analyzed by Chi-squared tests or Fisher's exact tests as appropriate. Continuous variables comparisons used Student's t-test (normal distribution) or Mann-Whitney U test (non-normal distribution). Wilcoxon signed-rank test analyzed paired ordinal data. Correlation analyses were calculated using Pearson's r and Spearman's ρ coefficients depending on data distribution. 5-year PFS and OS were assessed via Kaplan-Meier analysis with log-rank tests. Cox proportional hazards regression models were employed for univariate and multivariate survival analyses, presenting results as hazard ratios (HRs) with 95% confidence intervals (CIs). Data visualizations, including correlation clustering and survival curves, were implemented using the following R packages: ggplot2 (3.5.1), dplyr (1.1.4), survival (3.7–0), survminer (0.5.0), and pheatmap (1.0.12). Statistical significance was defined as p < 0.05. Abbreviations TMETumor immune microenvironmentLCISLung Cancer Immune ScoreLUADLung adenocarcinomaNSCLCNon-small cell lung cancerTAMsTumor-associated macrophagesNLRNeutrophil-to-lymphocyte ratioPLRPlatelet-to-lymphocyte ratioLDHLactate dehydrogenasePFSProgression-free survivalOSOverall survivalTIICTumor-infiltrating immune cellTMATissue microarray TMETumor immune microenvironmentLCISLung Cancer Immune ScoreLUADLung adenocarcinomaNSCLCNon-small cell lung cancerTAMsTumor-associated macrophagesNLRNeutrophil-to-lymphocyte ratioPLRPlatelet-to-lymphocyte ratioLDHLactate dehydrogenasePFSProgression-free survivalOSOverall survivalTIICTumor-infiltrating immune cellTMATissue microarray Ethics approval and consent to participate This study received ethical approval from the Medical Ethics Committee of Yan'an Hospital Affiliated to Kunming Medical University. Informed consent was obtained from all participants enrolled in the study. (Number: 2021– 140– 01). Funding This work was supported by the Scientific Research Fund Project of Yunnan Education Department (2024Y244, 2025Y0387), the Key Research and Development Program of Yunnan (202302AD080004, 202403AC100002), National Natural Science Foundation of China (32460155, 82260522). Clinical trials Not applicable. CRediT authorship contribution statement Wei Luo: Writing – original draft, Visualization, Investigation, Formal analysis. Jia Jia: Funding acquisition, Conceptualization. Qi Wang: Visualization, Formal analysis. Xiaoqing Li: Methodology, Formal analysis. Fang Yang: Investigation, Formal analysis. Yongtian Zhao: Formal analysis. Zilu Li: Formal analysis. Lin Li: Methodology, Data curation. Xiaodan Wang: Supervision, Project administration. Mingyao Meng: Supervision, Project administration. Haiyan Guo: Investigation. Wanli Zheng: Resources, Investigation. Liqiong Wang: Resources, Formal analysis. Dingbiao Li: Resources. Zongliu Hou: Supervision, Funding acquisition, Conceptualization. Yangfan Guo: Writing – review & editing, Funding acquisition, Conceptualization. Declaration of competing interest The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.