Abstract 918P
Background
The study was aimed to develop a deep learning model for predicting overall survival (OS) in lung cancer based on radiomic features, standard uptake value (SUV) and TNM staging system.
Methods
In this study patients’ outcome was predicted by a mixed model data generated through a machine learning approach, which integrated clinical variables and PET/CT imaging features from 320 NSCLC patients. A total of 48 potential predictors including 43 textural features, SUVmax, metabolic tumor volume (MTV) and TNM stage were used. Least absolute shrinkage and selection operator (LASSO) regression was used to select features with highest predictive value. Selected variables were given as input to a feed-forward neural multi-layer network (FNN). Mixed model was benchmarked against TNM-alone. Area under the curve (AUC) was used to assess diagnostic performance.
Results
Seventeen variables including 12 textural features, SUVmax, MTV and T, N, M stage showed highest predictive value at LASSO regression analysis. The AUC for outcome prediction with mixed model was 82.5% with 77.5%, 72% and 66% for accuracy, precision and F1 score respectively. AUC, accuracy, precision, and the F1 score obtained from TNM-only outcome prediction were 72%, 67%, 59% and 39% respectively. The Hanley test showed a significant difference between the AUCs obtained with the mixed model and the TNM alone outcome prediction (P<0.0001). Overall survival based on TNM prediction showed a significant difference from that observed and did not reach median survival (HR=0.32, P<0.0001). The difference between the observed and mixed model-predicted survival curves are not significant up to 22 months and become slightly significant starting from the 2nd year (HR=0.67, P<0.01).
Conclusions
These findings suggest that textural and PET imaging features are significant predictors of outcome in lung cancer. The association with tumor T, N and M stage outperforms TNM-alone and may serve as additional tool for patient management and stratification.
Clinical trial identification
Editorial acknowledgement
Legal entity responsible for the study
Dr Andrea Ciarmiello.
Funding
Has not received any funding.
Disclosure
All authors have declared no conflicts of interest.