Imaging AI prognosis of early stage lung cancer using CT radiomics

BackgroundAccurate prediction of lung cancer recurrence risk is crucial for treatment decisions and follow-up, particularly for Stage I patients who are not eligible for (neo-)adjuvant therapy but approximately one-third still recur after surgical resection. We present a machine learning model that uses patient computed tomography (CT) images and clinical features to predict lung cancer recurrence.MethodsA dataset of 968 clinical stage I-III lung cancer patients who underwent surgical resection was gathered from the US National Lung Screening Trial, the North Estonia Medical Centre, the Stanford University School of Medicine and Palo Alto Veterans Affairs Healthcare System. 313/968 had lung cancer recurrence, of which 221/776 were Stage 1. The pre-operative survival model was trained to predict the likelihood of recurrence at each time-point using radiomic features extracted from CT images and relevant clinical variables. An 8-fold cross-validation strategy was used and performance evaluated using the time-dependent Area-Under-the-ROC-Curve (AUC), disease-free survival (DFS), hazard ratio (HR) and log-rank test against clinical staging alone.ResultsThe ML survival model was better able to stratify patients into high and low-risk (HR=2.7, p<0.005) compared with Stage I vs II-III (HR=2.2, p<0.005). The same was observed for the Stage I sub-group (HR=2.4, p<0.005) when compared with using Stage IA vs IB (HR=1.1, p=0.79). The gaps between the high and low-risk DFS at 1, 2, and 5 years are larger for the ML model than separation by staging. The ML survival model also had better prediction accuracy, with the time-dependent AUCs being significantly better than staging at 1, 2, and 5-year marks. ML model threshold was set to match on high/low-risk patient counts. Table: 1214P Predictors 1-year DFS % low-risk 1-year DFS % high risk 2-year DFS % low risk 2-year DFS % high-risk 5-year DFS % low risk 5-year DFS % high risk HR 1-year AUC 2-year AUC 5-year AUC cTNM - Stage I (n=776) vs Stage II-III (n=192) 92.0 77.1 85.6 66.7 74.0 50.8 2.2, (1.7, 2.8) (p<0.005) 0.66 0.64 0.61 ML model - low-risk (n=776) vs high-risk (n=192) 93.2 72.6 86.8 61.9 75.3 46.2 2.7, (2.1, 3.4) (p<0.005) 0.74 0.70 0.68 cTNM - Stage IA (n=116) vs Stage IB (n=660) 92.2 91.0 86.1 82.9 74.4 74.0 1.1, (0.7, 1.5) (p=0.79) 0.52 0.53 0.50 ML model - Stage I low-risk (n=116) vs Stage I high-risk (n=660) 94.0 80.9 88.0 72.0 77.4 55.1 2.4, (1.8, 3.3) (p<0.005) 0.67 0.64 0.64 ConclusionsThe ML survival model outperforms clinical staging in patient risk-stratification and time-dependent lung cancer recurrence prediction. With further development, this algorithm could prove a valuable tool to aid the management of lung cancer patients.Legal entity responsible for the studyThe authors.FundingOptellum.DisclosureA. Gasimova, B. Heames, N. Waterfield Price: Financial Interests, Institutional, Other, Employee: Optellum. G. Hodgkinson: Financial Interests, Institutional, Other, Employee: Medtronic. L. Freitag: Financial Interests, Institutional, Advisory Board: Optellum. All other authors have declared no conflicts of interest.