Late-stage diagnosis drives lung cancer mortality in symptomatic patients. Lung cancer remains the leading cause of cancer-related deaths worldwide, with 48% of cases diagnosed at the distant metastasis stage where the 5-year survival rate is approximately 8%. Symptomatic patients face compounding delays: one-third return to their GP three or more times with lung cancer-associated symptoms before specialist referral, and the UK National Cancer Diagnosis Audit documented primary care delays of 60 to 90 days for 17.9% and 10.8% of suspected lung cancer patients, respectively.
Existing screening tools have critical gaps that minimally invasive tests could address. Low-dose CT screening reduces mortality by approximately 20% but is limited to high-risk smokers and excludes the roughly 25% of lung cancer patients who are never-smokers. ctDNA-based multicancer early detection tests such as Galleri achieve specificity near 99% but sensitivity as low as 8.7% for stage I lung cancer in the symptomatic setting, and 0% in the asymptomatic setting. A minimally invasive blood test with substantially higher sensitivity would complement or reduce reliance on CT imaging.
Peripheral immune cell changes reflect systemic cancer surveillance activity. Cancer-associated immune surveillance is reflected in circulating peripheral blood mononuclear cells (PBMCs), with changes in immune cell subset frequencies and functional states providing an indirect window into tumor-immune interactions. Single-cell RNA sequencing and CyTOF have identified circulating anti-tumor CD8 T cells and monocyte subsets that predict immunotherapy response, and cancer-specific T cell receptor repertoires have been shown to detect early-stage lung cancer in LDCT-screened cohorts, confirming that immune biomarkers can operate before tumors shed detectable nucleic acids.
Automated CT analysis reduces radiologist workload while adding complementary information. The diagnostic pathway for lung cancer typically requires radiologist assessment at multiple steps including chest X-ray, CT, PET-CT, and biopsy guidance. Radiologist shortages across the UK and Europe are creating examination delays that worsen diagnostic timelines. Automated image analysis through CT texture analysis (radiomics) and deep learning autoencoders can extract quantitative tumor features from CT scans that complement subjective radiologist assessment and could reduce manual workload.
344 symptomatic patients were recruited prospectively from a lung cancer clinic. The LungExoDETECT study enrolled patients referred to secondary care with clinical symptoms or signs suspicious of lung cancer at the Lister Hospital, Hertford County, UK between October 2020 and November 2021. Peripheral blood samples and standard-of-care CT scans were collected simultaneously. After excluding 174 patients due to absent or sub-6mm lesions or non-lung cancer diagnoses, 170 patients were included in the analysis set, of whom 79 (46.5%) received a lung cancer diagnosis and 23 (29.1%) were diagnosed at stage I.
Four complementary data modalities generated 187 total candidate features. CT scan data were analyzed using CT Texture Analysis (CTTA) generating 42 features from filtration-histogram-based texture analysis, and Deep Learning Autoencoder (DLA) generating 32 latent space features from an unsupervised representation of lesion image characteristics. Peripheral blood was analyzed by high-depth flow cytometry generating 45 immune cell features, and plasma exosome dot-blot analysis generating 68 protein component features. Each modality was analyzed independently before multi-modal fusion.
Bayesian multivariate regression performed covariate selection and model building simultaneously. Bayesian multivariate regression (BMR) was selected over LASSO or Elastic Net for its ability to optimally balance the number of selected covariates against the information content of the training data, inherently preventing overfitting without requiring separate regularization parameter tuning. The cohort was split 75/25 into training and test sets, with random splitting repeated eight times. Only covariates selected in at least 50% of the eight independent training runs were retained in the final signature, ensuring that selected features were robust to patient assignment variation.
Data fusion was tested at three levels to optimize the combined signature. Multi-modal combination was evaluated through early fusion (combining raw covariates from all modalities before modeling), intermediate fusion (combining features selected by individual modality models), and late fusion (combining the risk scores output by individual modality signatures). The three fusion approaches were compared on both training and test sets to identify the strategy that maximized predictive performance while avoiding overfitting.
Deep learning autoencoder outperformed CT texture analysis as an imaging modality. On training sets, DLA achieved AUC 0.72-0.77 compared to CTTA AUC 0.67-0.73. On test sets, DLA AUC ranged 0.56-0.71 while CTTA ranged 0.57-0.79, with overlapping ranges reflecting individual split variability. The two CTTA features consistently selected by BMR were mean CT attenuation at spatial scale filter 0 (reflecting tumor density) and mean positive pixel intensity at spatial scale filter 3 (reflecting average brightness at a 3mm fine texture scale), both interpretable as direct physical tissue property measurements.
Flow cytometry immune signature substantially outperformed exosome protein analysis. The flow cytometry immune signature achieved training AUC 0.66-0.70, whereas the exosome dot-blot signature achieved only AUC 0.50-0.60 on training sets, with inconsistent feature selection across repeated runs. Exosome-based features were therefore excluded from combined model development. The immune signature was driven by two highly consistent biomarkers: elevated KIR3DL1-expressing CD8 T lymphocytes indicating lung cancer, and elevated type 2 dendritic cells (cDC2) indicating non-cancer pathology.
Combining immune blood data with CT imaging improved overall AUC to 0.81. The combined immune plus DLA signature achieved training AUC 0.77-0.83 and test AUC 0.68-0.81. The combined immune plus CTTA signature achieved training AUC 0.79-0.84 and test AUC 0.66-0.82. The final late fusion signature across all three modalities achieved ROC AUC 0.81 with sensitivity 0.72 and specificity 0.77 at the Youden optimal threshold, compared to AUC 0.69 for immune alone, 0.70 for CTTA alone, and 0.73 for DLA alone as single modalities.
Stage-consistent performance was observed including for stage I cancer detection. The combined signature achieved sensitivity of 76-78% across most cancer stages, confirming that performance was not driven solely by late-stage tumors with larger lesions. The 23 stage I cancers in this symptomatic cohort provided a meaningful subgroup for early-stage evaluation. The immune signature alone detected 82% of 11 non-lung cancers in the cohort, further demonstrating that the immune profiling component captures a general cancer-related immune shift beyond lung-specific signals.
KIR3DL1-expressing CD8 T cells were the strongest single predictor of lung cancer. KIR3DL1 is a member of the Killer Cell Inhibitory Receptor (KIR) family, whose expression on CD8 T cells has been linked to effector cells with reduced proliferative capacity and impaired IFN-gamma production following T cell receptor engagement. In the LungExoDETECT cohort, the KIR3DL1+ CD8 population also exhibited low expression of CD107a (Lamp1), indicating that these cells were not engaged in degranulation and were functionally exhausted rather than actively cytotoxic.
KIR expression on CD8 T cells reflects chronic tumor antigen stimulation. KIR expression in CD8 T cells has been linked to epigenetic demethylation following chronic TCR stimulation, which is consistent with the progressive T cell exhaustion that accompanies cancer immunosurveillance. The inhibitory KIR-HLA interaction does not directly impair T cell degranulation but compromises activation-induced transcription required for clonal expansion and cytokine production, contributing to the functional immunosuppression observed in tumor-bearing patients.
Elevated cDC2 distinguished non-cancer pathology including infection and allergy. In non-cancer patients, whose primary diagnoses included infection and allergic reactions, type 2 conventional dendritic cells (cDC2) were consistently elevated. cDC2 cells play a central role in antigen presentation for TH2 helper responses governing humoral immunity and have been correlated with TH17 activation in chronic allergy. The cDC2 cells in this cohort expressed high HLADR and CD38 indicating active pro-inflammatory function, and did not express PDL2, confirming they were not functionally suppressed.
Both immune features were selected in all eight independent training runs. The two immune biomarkers, elevated KIR3DL1+ CD8 T cells and reduced cDC2 relative frequency, were retained in the final model in the majority of all eight independent randomized training runs, providing strong evidence for their biological relevance and statistical robustness independent of which patients were assigned to training versus test sets. This consistency of feature selection is the primary evidence for the reliability of the immune signature.
Bayesian regression outperformed Elastic Net for test set generalization. An Elastic Net algorithm comparison using the same feature sets produced stronger AUC values on training sets but substantially worse test set performance due to overfitting. Bayesian multivariate regression inherently selects the optimal number of covariates based on information content in the training data, rather than applying external regularization parameters that require independent tuning. This property makes BMR models more generalizable to new patient populations without overfitting to idiosyncratic training set characteristics.
Deep learning autoencoder latent space features are interpretable through image reconstruction. A preliminary latent space interpretation was performed by manually varying individual DLA feature values and observing the decoded image output. The five BMR-selected DLA features were found to correspond to lesion size, morphology (including spiculation and sphericity), location within the lung (evidenced by pleura distance and orientation changes), and degree of pleural attachment. These image properties are directly clinically meaningful, addressing the black-box criticism of deep learning models for clinical applications.
The 6mm lesion size cutoff may limit early-detection performance. Current clinical practice excludes management action for lesions smaller than 6mm, which was applied as an exclusion criterion for this study. However, the authors acknowledge that relaxing this cutoff in future studies could improve sensitivity for the very earliest cancers, at the cost of requiring longer follow-up to obtain definitive diagnoses and a different benchmarking framework than the standard NHS National Optimal Lung Cancer Pathway used here.
Future improvements include retraining the autoencoder on matched clinical data. The DLA was pretrained on the LUNA16 public dataset rather than on the study-acquired images, creating a potential domain mismatch between the training domain and the clinical images. Retraining the deep learning autoencoder on a dataset more closely matched to the clinical scanning parameters and patient population is identified as the most direct path to further improving the imaging component of the combined signature.
Combining immune monitoring with CT imaging effectively improves diagnostic accuracy. This study demonstrated that a blood-based immune signature derived from peripheral flow cytometry, when combined with automated CT analysis, achieves an AUC of 0.81 with sensitivity 0.72 and specificity 0.77 in symptomatic patients with suspected lung cancer. This performance is comparable to multicancer early detection tests currently being adopted in clinical practice, while offering substantially higher early-stage sensitivity than ctDNA-based tests.
Immune-based biomarkers detect cancer before nucleic acid shedding becomes detectable. A fundamental advantage of immune cell-based detection is that systemic immune changes in response to tumor antigens occur when tumors are relatively small, at an earlier point in tumor development than when sufficient nucleic acid fragments enter circulation to be detected by ctDNA or cell-free RNA tests. The 82% sensitivity of the immune signature alone for detecting non-lung cancers in the cohort further demonstrates that the immune response to malignancy is broadly detectable across cancer types.
Prospective validation in an independent cohort is the essential next step. While the study used eight repeated randomizations and strict 25% holdout test sets within the 170-patient cohort to estimate performance, a completely independent external validation cohort with the same depth of immune and imaging profiling is required before clinical implementation. Future work should also explore incorporating metabolomic parameters that could strengthen immune observations through causative correlations, and investigate whether reducing the lesion size threshold below 6mm improves sensitivity for the earliest detectable cancers.