[¹⁸F]fluorodeoxyglucose positron emission tomography/computed tomography in combination with clinical data in...

Positron emission tomography/computed tomography (PET/CT) has an established role in evaluating patients with lung cancer. The aim of this work was to assess the predictive capability of [¹⁸F]Fluorodeoxyglucose ([¹⁸F]FDG) PET/CT parameters on overall survival (OS) in lung cancer patients using an artificial neural network (ANN) in parallel with conventional statistical analysis.

Retrospective analysis was performed on a group of 165 lung cancer patients (98M, 67F). PET features associated with the primary tumor: maximum and mean standardized uptake value (SUV_max, SUV_mean), total lesion glycolysis (TLG) metabolic tumor volume (MTV) and area under the curve-cumulative SUV histogram (AUC-CSH) and metastatic lesions (SUV_maxtotal, SUV_meantotal, TLG_total, and MTV_total) were evaluated. In parallel with conventional statistical analysis (Chi-Square analysis for nominal data, Student's t test for continuous data), the data was evaluated using an ANN. There were 97 input variables in 165 patients using a binary classification of either below, or greater than/equal to median survival post primary diagnosis. Additionally, phantom study was performed to assess the most optimal contouring method.

Males had statistically higher SUV_max (mean: 10.7 vs 8.9; p = 0.020), MTV (mean: 66.5 cm³ vs. 21.5 cm³; p = 0.001), TLG (mean 404.7 vs. 115.0; p = 0.003), TLG_total (mean: 946.7 vs. 433.3; p = 0.014) and MTV_total (mean: 242.0 cm³ vs. 103.7 cm³; p = 0.027) than females. The ANN after training and validation was optimised with a final architecture of 4 scaling layer inputs (TLG_total, SUV_maxtotal, SUV_meantotal and disease stage) and receiving operator characteristic (ROC) analysis demonstrated an AUC of 0.764 (sensitivity of 92.3%, specificity of 57.1%).

Conventional statistical analysis and the ANN provided concordant findings in relation to variables that predict decreased survival. The ANN provided a weighted algorithm of the 4 key features to predict decreased survival.

Identification of parameters which can predict survival in lung cancer patients might be helpful in choosing the group of patients who require closer look during the follow-up.