Fusing Data from CT Deep Learning, CT Radiomics and Peripheral Blood Immune profiles to Diagnose Lung Cancer in Symptomatic Patients

Background

Lung cancer is the leading cause of cancer-related deaths. Diagnosis at late stages is common due to the largely non-specific nature of presenting symptoms contributing to high mortality. There is a lack of specific, minimally invasive low-cost tests to screen patients ahead of the diagnostic biopsy.

Patients and Methods

344 symptomatic patients from the lung clinic of Lister hospital suspected of lung cancer were recruited. Predictive covariates were successfully generated on 170 patients from Computed Tomography (CT) scans using CT Texture Analysis (CTTA) and Deep Learning Autoencoders (DLA) as well as from peripheral blood data for immunity using high depth flow-cytometry and for exosome protein components. Predictive signatures were formed by combining covariates using Bayesian regression on a randomly chosen 128-patient training set and validated on a 42-patient held-out set. Final signatures were generated by fusing the data sources at different levels.

Results

Immune and DLA were the best single modality signatures with test set AUCs of 0.76 (95% CI: 0.61 – 0.91) and 0.75 (95% CI: 0.60 - 0.90) respectively. The final combined signature had a ROC AUC of 0.86 (95% CI: 0.73 - 0.99) on the withheld test set. The overall sensitivity and specificity were 0.722 and 0.901 respectively.

Conclusions

Combining immune monitoring with CT scan data is an effective approach to improving sensitivity and specificity of Lung cancer screening even in symptomatic patients.

Highlights

• Combining immune monitoring in peripheral blood with CT scan data improves lung cancer screening sensitivity and specificity.

• Elevated levels of KIR3DL1+ CD8 T cells may be indicative of cancer.

• A cancer biomarker that combines a deep learning autoencoder with peripheral immune profiling achieved a 0.86 ORC AUC.