AI Radiomics for the Early Detection of Lung Cancer

20 Nov 2026 11:00 11:15
Ren Yuan Speaker Canada
AI Radiomics for the Early Detection of Lung Cancer Background: Lung cancer remains the leading cause of cancer-related mortality in Canada. Low-dose computed tomography (LDCT) screening frequently detects small or ground-glass opacity nodules that are challenging to assess for malignancy. This study aimed to develop a radiomics-based machine-learning model to predict the future malignant behaviour of these difficult nodules. Materials and Methods: Small nodules were drawn from three lung cancer screening datasets: PanCan (n=2,085 participants; N=10,381 nodules), ILST (n=1,082 participants; N=2,389 nodules), and DLCS (n=1,373 participants; N=2,009 nodules). A self-attention deep-learning model was trained for nodule segmentation, and axial masks were reconstructed into three-dimensional core nodule masks. Radiomics features were extracted from four discrete regions: the nodule core and surrounding regions, including core-plus-edge, shell, and shell-without-blood-vessel masks. Machine-learning models were developed and compared with the Brock model. Results: At a low false-positive rate of 10%, the ML-radiomics models demonstrated superior sensitivity to the Brock model across all nodule types in the external validation set. Both approaches identified malignant nodules at baseline that would not have otherwise received interventional management and predicted malignancy up to 2 years before clinical confirmation. Conclusion: A pipeline integrating deep-learning segmentation with ML-radiomics can accurately classify subcentimetre nodules up to approximately 2 years before clinical confirmation of malignancy in lung cancer screening. Keywords: Lung Cancer, LDCT Screening, Radiomics, Machine Learning, Deep Learning