Deep Learning Features and Metabolic Tumor Volume Based on PET/CT to Construct Risk Stratification in Non-small Cell Lung Cancer

Abstract

Rationale and objectives: To build a risk stratification by incorporating PET/CT-based deep learning features and whole-body metabolic tumor volume (MTV_wb), which was to make predictions about overall survival (OS) and progression-free survival (PFS) for those with non-small cell lung cancer (NSCLC) as a complement to the TNM staging.

Materials and methods: The study enrolled 590 patients with NSCLC (413 for training and 177 for testing). Features were extracted by employing a convolutional neural network. The combined risk stratification (CRS) was constructed by the selected features and MTV_wb, which were contrasted and integrated with TNM staging. In the testing set, those were verified.

Results: Multivariate analysis revealed that CRS was an independent predictor of OS and PFS. C-indexes of the CRS demonstrated statistically significant increases in comparison to TNM staging, excepting predicting OS in the testing set (for OS, C-index=0.71 vs. 0.691 in the training set and 0.73 vs. 0.736 in the testing set; for PFS, C-index=0.702 vs. 0.686 in the training set and 0.732 vs. 0.71 in the testing set). The nomogram that combined CRS with TNM staging demonstrated the most superior model performance in the training and testing sets (C-index=0.741 and 0.771).

Conclusion: The addition of CRS improves TNM staging's predictive power and shows potential as a useful tool to support physicians in making treatment decisions.

Keywords: Deep learning; Non-small cell lung cancer; Overall survival; Positron emission tomography/computed tomography; Risk stratification.