The problem with standard imaging after immunotherapy. When patients with locally advanced non-small cell lung cancer (NSCLC) receive neoadjuvant immunochemotherapy before surgery, their tumors undergo complex biological changes - not just shrinkage, but tissue remodeling that creates mixtures of viable tumor cells, necrosis, fibrosis, and inflammatory infiltrates. Standard CT-based response criteria (RECIST) were designed for chemotherapy and perform poorly in this setting, producing unreliable predictions of actual pathological response.
Tumor-infiltrating lymphocytes as the key indicator. Tumor-infiltrating lymphocytes (TILs) - cancer-fighting immune cells that have entered the tumor - are among the strongest predictors of response to immunotherapy. Patients with high TIL infiltration achieve far higher rates of major pathological response (residual tumor of 10% or less). However, measuring TIL levels currently requires surgical biopsy and histological staining, limiting its use for monitoring patients during treatment before surgery.
Habitat radiomics as a non-invasive alternative. Rather than treating a tumor as a single uniform region, habitat radiomics uses unsupervised clustering to identify distinct sub-regions within the tumor that have different imaging characteristics - corresponding to biologically different zones such as active tumor, necrosis, and fibrosis. This study developed and validated a habitat radiomics model capable of predicting TIL status from post-treatment CT images without tissue sampling.
Combined imaging and single-cell biology. Beyond the radiomics model, the study used public single-cell RNA sequencing (scRNA-seq) data to identify the molecular mechanism behind treatment resistance. This revealed a specific immunosuppressive regulatory T-cell subpopulation that accumulates in non-responding tumors and that habitat radiomics may indirectly reflect through its imaging phenotype.
Clinical cohort and patient selection. The retrospective study enrolled 238 NSCLC patients (stages IB-IIIB) who received at least two cycles of neoadjuvant immunochemotherapy followed by surgical resection at Union Hospital between August 2019 and February 2024. The most common histological subtype was squamous cell carcinoma (66.8%), with a median age of 62 years and male predominance (89.4%). Of 238 patients, 201 met all criteria for radiomics analysis, including adequate CT image quality and CT acquired within 30 days of surgery.
TIL assessment and pathological response classification. Surgical specimens were assessed by pathologists blinded to clinical outcomes. TIL-positive status was defined as lymphocyte infiltration of at least 10% of the stromal area. Major pathological response (MPR) was defined as residual viable tumor of 10% or less. In total, 157 patients (66%) were TIL-positive, 129 (54.2%) achieved MPR, and 79 (33.1%) achieved pathological complete response (pCR).
Habitat identification using K-means clustering. Post-treatment contrast-enhanced CT images were segmented manually, then voxel-level radiomic features were extracted using PyRadiomics for each voxel within the tumor. K-means clustering was applied to identify distinct imaging sub-regions (habitats). The optimal number of clusters was determined using the Calinski-Harabasz index, which identified three habitat sub-regions as optimal across all patients.
Feature selection and model development. Initial extraction produced 5,505 features from three habitat sub-regions. After Mann-Whitney U filtering (retaining 984 features), Pearson correlation filtering (retaining 137), and LASSO regression (selecting 24 final features), seven machine learning algorithms were trained and compared: Logistic Regression, Random Forest, XGBoost, LightGBM, ExtraTrees, Multi-layer Perceptron, and GradientBoost. A 70/30 training-test split was used with five-fold cross-validation for robustness assessment.
Random Forest outperforms other algorithms. Among seven models evaluated, Random Forest (RF) achieved the best balance of sensitivity and specificity: AUC 0.823 (95% CI: 0.694-0.932) in the independent test set, with sensitivity of 0.786, specificity of 0.737, and F1 score of 0.825. While XGBoost achieved marginally higher AUC of 0.845, its sensitivity of 0.762 and specificity of 0.632 were less balanced, making RF the preferred model for clinical use.
Habitat radiomics substantially outperforms whole-tumor radiomics. A comparison model built on whole-tumor radiomic features (without habitat sub-region analysis) achieved only AUC 0.679 - compared to 0.823 for the habitat model. This 21% relative improvement in AUC confirms that capturing spatial heterogeneity through habitat analysis adds substantial predictive information beyond treating the tumor as a uniform mass.
TIL status strongly predicts treatment outcomes. The association between TIL status and pathological response was striking: among TIL-positive patients, 70.1% achieved MPR and 45.9% achieved pCR. Among TIL-negative patients, only 23.5% achieved MPR and 8.6% achieved pCR. TIL-positive status also predicted significantly improved recurrence-free survival (RFS), including in the non-pCR subgroup - where TIL-positive patients still showed meaningfully better outcomes than TIL-negative patients (p=0.041).
Radiomics model stratifies recurrence risk. Using the Youden-optimal threshold, the RF model stratified patients into high-risk and low-risk groups based on predicted TIL status. Low-risk patients showed significantly improved RFS compared to high-risk patients (p=0.018). Traditional clinical staging variables (T stage, N stage, clinical stage, smoking status, lesion location) did not achieve prognostic stratification in this immunotherapy-treated population, confirming that imaging-predicted TIL status carries prognostic information that standard staging does not.
Why tumor spatial heterogeneity matters. Standard radiomics extracts features from the entire tumor volume as a single entity, averaging over regions that may be biologically very different. Active tumor tissue, necrotic zones, and fibrotic regions each have distinct cellular compositions, vascular patterns, and metabolic states - and therefore different CT imaging characteristics. Treating these as one entity discards information about their spatial arrangement, which may reflect immune status.
The three habitat sub-regions identified. K-means clustering consistently identified three distinct habitat sub-regions within post-treatment tumors, confirmed by the Calinski-Harabasz index peak at k=3 (score 393,162, with a sharp decline to 363,465 at k=4). Feature heatmaps showed heterogeneous distributions across these sub-regions, with warmer colors in some zones and cooler in others - reflecting real biological differences captured by the radiomic feature values within each habitat.
SHAP analysis reveals which features drive predictions. SHapley Additive exPlanations (SHAP) analysis identified the specific radiomic features contributing most to TIL predictions in the RF, XGBoost, and LightGBM models. This interpretability step links imaging texture characteristics within specific habitat sub-regions to the predicted immune status, providing a pathway to understand which CT texture patterns are associated with immune infiltration.
Biological validation with immunofluorescence. To confirm that the imaging habitats reflected real immune differences, multiplex immunofluorescence staining was performed on representative tissue sections using CD3 (T cells), CD20 (B cells), and FOXP3 (regulatory T cells). MPR tumors showed high CD3+ T-cell infiltration with minimal FOXP3+ Tregs - an active anti-tumor immune state. Non-MPR tumors showed the reverse: elevated FOXP3+ Treg infiltration, confirming the imaging-based prediction of immunosuppressive biology.
Single-cell profiling reveals the immunosuppressive architecture. Analysis of a public scRNA-seq dataset from NSCLC patients who received neoadjuvant immunochemotherapy (GEO: GSE207422, 17,533 cells) revealed a compositional shift in non-responding tumors. Non-MPR tumors showed lower T-cell proportions overall, with relative expansion of the myeloid compartment - an immune-excluded architecture. Within the T-cell compartment, non-MPR tumors showed simultaneous expansion of both regulatory T cells (Tregs) and exhausted CD8+ T cells, indicating a dual immunosuppressive state.
Naive T cells differentiate into Tregs in resistant tumors. Pseudotime trajectory analysis of the CD4+ T-cell lineage reconstructed a continuous differentiation pathway from naive T cells to a Treg fate. As cells progressed along this trajectory, the naive T-cell signature (TCF7, CCR7) was replaced by FOXP3 and co-inhibitory receptor expression characteristic of Tregs. This demonstrates that the Treg accumulation in non-responding tumors is not simply due to recruitment of pre-existing Tregs, but to active differentiation of naive T cells into the suppressive Treg fate.
SERPINB9+ Tregs as the key suppressive population. A specific Treg subpopulation distinguished by high SERPINB9 expression was significantly more abundant in non-MPR patients. SERPINB9 is a stoichiometric inhibitor of Granzyme B - the primary cytotoxic molecule used by killer T cells and NK cells to destroy cancer cells. By producing SERPINB9, these Tregs protect themselves and potentially surrounding tumor cells from cytotoxic killing, enabling their persistence within the inflamed tumor core.
Divergent cell communication networks in responders versus non-responders. CellChat analysis of intercellular signaling revealed opposing architectures. In responding tumors, communication was dominated by MHC-I and MHC-II antigen presentation pathways and TNF-lymphotoxin signaling - pathways that support productive adaptive immunity and tertiary lymphoid structure formation. In non-responding tumors, the network was rewired toward immunosuppression: upregulation of the Galectin-9 and TIM-3 checkpoint axis, PGE2 signaling, and VCAM1/collagen pathways promoting fibrotic stromal remodeling, combined with dominant CD137 and OX40 signaling that promoted Treg proliferation rather than effector T cell activation.
Blood-based inflammatory markers as complementary predictors. Multivariate logistic regression of clinical variables identified pre-treatment neutrophil-to-lymphocyte ratio (NLR) and post-treatment platelet-to-lymphocyte ratio (PLR) as independent predictors of MPR achievement. NLR reflects systemic immune activation and baseline inflammatory state, while PLR reflects post-treatment immune dynamics. These simple blood tests complement imaging-based TIL prediction and could be integrated with the radiomics model in future multi-modal prediction tools.
Proposed clinical applications. A validated imaging-based TIL predictor could serve several practical roles: identifying patients who are unlikely to achieve adequate response and may benefit from treatment intensification (additional cycles or dose escalation) before surgery; selecting candidates for adjuvant immunotherapy after surgical resection; and monitoring treatment response dynamically without repeated invasive sampling. The 4-second per scan inference time makes integration into clinical workflows feasible.
Therapeutic implications of SERPINB9+ Tregs. The identification of SERPINB9+ Tregs as a driver of resistance points toward potential treatment strategies. Anti-SerpinB9 therapy has been shown to achieve tumor killing and enhance immunotherapy in experimental settings. OX40/CD137 bispecific agonist therapy has been shown to reprogram Tregs, potentially reversing the immunosuppressive state. These findings provide preclinical rationale for combination strategies targeting this specific Treg population in immunotherapy-resistant NSCLC.
Limitations of the current study. External multi-center validation is the most important gap - model performance at other institutions, with different scanner types and patient populations, remains to be established. The scRNA-seq analysis used a public dataset with an imbalanced MPR/non-MPR group ratio (3 vs. 10 patients), which limits the generalizability of the cellular mechanism findings. TIL assessment used semi-quantitative pathological evaluation rather than whole-slide digital analysis, which may introduce inter-observer variation. Binary TIL classification may also not fully capture the spectrum of immune infiltration patterns.