Fusing Data from CT Deep Learning, CT Radiomics and Peripheral Blood Immune profiles to Diagnose Lung Cancer in a Cohort of Patients Experiencing Symptoms

Technical Summary: Fusing CT Deep Learning, CT Radiomics, and Peripheral Blood Immune Profiles for Lung Cancer Diagnosis

Problem Statement
Lung cancer remains the leading cause of cancer-related deaths globally, primarily due to late-stage diagnosis. Current screening methods, such as low-dose computed tomography (LDCT), are limited by high false-positive rates, exclusion of non-smokers and younger populations, and significant radiologist workload leading to examination delays. Furthermore, existing liquid biopsy approaches based on circulating tumor DNA (ctDNA) have shown low sensitivity for early-stage (Stage I) lung cancer in symptomatic settings. There is a critical need for a minimally invasive, robust diagnostic tool that can improve sensitivity and specificity for patients presenting with symptoms, potentially reducing reliance on manual radiologist interpretation and detecting malignancies earlier than nucleic acid-based shedding allows.

Methodology
The study, titled "LungExoDETECT," was a prospective cohort study conducted between October 2020 and November 2021 involving 344 patients referred to lung clinics in the UK with symptoms suspicious of lung cancer.

• Data Acquisition:
  • Imaging: Patients underwent standard CT scans. Lesions were analyzed using two distinct computational approaches:
    1. CT Texture Analysis (CTTA): Extracting 42 features based on filtration-histogram methods (using TexRAD software).
    2. Deep Learning Autoencoders (DLA): Utilizing a deep learning model to generate 32 latent vector features representing lesion characteristics (e.g., size, morphology, pleural attachment).
  • Blood Analysis: Peripheral blood samples were collected and analyzed via:
    1. High-depth Flow Cytometry: Profiling 45 immune cell subsets.
    2. Exosome Protein Dot-blot: Analyzing 68 exosome protein components.
• Model Construction:
  • From the initial 344 patients, 174 were excluded (lesions <6mm, no visible lesion, or non-lung cancer diagnosis), leaving 170 patients (79 lung cancer, 91 non-cancer) for analysis.
  • The cohort was split into a training set (75%) and a held-out test set (25%). This split was repeated 8 times via bootstrapping to ensure robustness.
  • Bayesian Multi-variate Regression (BMR) was employed for signature generation. Unlike regularized methods (e.g., LASSO/Elastic Net) which can overfit, BMR was chosen for its ability to perform automatic covariate selection and provide interpretable weights while minimizing overfitting.
  • Fusion Strategies: The study evaluated three levels of data fusion:
    1. Early Fusion: Combining raw covariates from all modalities.
    2. Intermediate Fusion: Combining features selected from individual signatures.
    3. Late Fusion: Combining the output risk scores of individual signatures.

Key Contributions

1. Novel Biomarker Integration: The study proposes a diagnostic signature that fuses automated CT image analysis (both radiomics and deep learning) with peripheral immune profiling, specifically focusing on immune response rather than nucleic acid shedding.
2. Identification of Specific Immune Signatures: The analysis identified two robust immune predictors:
   • Upregulation of KIR3DL1+ CD8 T lymphocytes: Associated with lung cancer presence. These cells exhibited an exhausted functional state (low CD107a expression).
   • Downregulation of Type 2 Dendritic Cells (cDC2): Associated with non-cancer pathologies (often infection or allergic reactions).
3. Interpretability: The use of Bayesian regression allowed for a "white box" model where specific biological components (e.g., specific immune cell subsets and CT texture scales) contributing to the prediction were explicitly identified and weighted.
4. Validation in Symptomatic Cohort: The model was validated specifically in patients already experiencing symptoms, a setting where early detection is most challenging and clinically urgent.

Results

• Single Modality Performance:
  • DLA (Deep Learning): Achieved the highest single-modality AUC (0.73), outperforming CTTA (0.70) and the immune signature (0.69).
  • Immune Signature: Performed significantly better than the exosome protein signature, which was inconsistent and not used in the final combined model.
• Combined Performance:
  • The final combined signature, fusing immune data with CT data (specifically the DLA and CTTA models), achieved a ROC AUC of 0.81.
  • Sensitivity: 0.72 (72%) overall. Sensitivity remained high (76–78%) across most stages, with a noted drop in Stage II (attributed to small sample size in that subgroup).
  • Specificity: 0.77 (77%), which is notable given the cohort consisted of patients with advanced symptoms and non-cancer pathologies.
• Non-Lung Cancer Detection: When applied to 11 patients with non-lung cancers (using only the immune signature as CT was irrelevant), the model detected 9/11 cases (82% sensitivity), missing only two breast cancer cases.
• Feature Selection: The final model consistently selected specific covariates across the 8 bootstrapped runs, including the immune markers (KIR3DL1+ CD8 and cDC2) and specific CT features (Mean intensity at SSF=0 and mean intensity of positive pixels at SSF=3).

Significance and Claims
The paper claims that combining immune monitoring with automated CT scan analysis offers a viable route to improving the sensitivity and specificity of lung cancer screening, even in symptomatic patients. The authors emphasize that:

• Immune Response as an Early Marker: Immune cell-based detection may occur at an earlier time point than the shedding of nucleic acids (ctDNA), potentially addressing the low sensitivity of current liquid biopsies for early-stage disease.
• Reduced Radiologist Workload: By utilizing automated deep learning and radiomics, the approach aims to alleviate the burden on radiologists, which is a known bottleneck in current screening pathways.
• Clinical Utility: The resulting mathematical model is highly interpretable, pinpointing specific immune system components that drive the prediction. The authors note that while the AUC of 0.81 is within the realm of clinical utility, the model was developed for symptomatic patients and further validation in independent cohorts and different screening populations is required.
• Methodological Robustness: The study highlights the advantage of Bayesian regression over Elastic Net in this context, citing its superior ability to prevent overfitting and select an optimal number of covariates based on evidence rather than just regularization.

The authors conclude that this multi-modal approach reinforces the value of systemic immune response profiling as a complementary tool to imaging for the early detection of lung cancer.